Inhibition stabilized network model in
the primary visual cortex
Studies on conditions to achieve surround suppression and
properties of spontaneous and sensory-driven activities
Jun Zhao
Submitted in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy
in the Graduate School of Arts and Sciences
COLUMBIA UNIVERSITY
2012
Abstract
Inhibition stabilized network model in the primary visual cortex
Jun Zhao
In this paper, we studied neural networks of both excitatory and inhibitory populations with
inhibition stabilized network (ISN) models. In ISN models, the recurrent excitatory connections
are so strong that the excitatory sub-network is unstable if the inhibitory firing rate is fixed;
however, the entire network is stable due to inhibitory connections. In such networks, external
input to inhibitory neurons reduced their responses due to the withdrawal of network excitation
(Tsodyks et al., 1997). This paradoxical effect of the ISN was observed in recent surround
suppression experiments in the primary visual cortex with direct membrane conductance
measurements (Ozeki et al., 2009). In our work, we used a linearized rate model of both
excitatory and inhibitory populations with weight matrices dependent on the locations of the
neurons. We applied this model to study surround suppression effects and searched for networks
with appropriated parameters. The same model was also applied in the study of spontaneous
activities in awake ferrets. Both studies led to network solutions in the ISN regime, suggesting
that ISN mechanisms might play an important role in the neural circuitry in the primary visual
cortex.
~ i ~
Table of Contents
Table of Contents ............................................................................................................................. i
List of Figures ................................................................................................................................ iv
Acknowledgements ........................................................................................................................ vi
Chapter I: Introduction and Literature Review ............................................................................... 1
1. Surround suppression effects in the primary visual cortex (V1) .......................................... 1
2. Spontaneous and sensory-driven activities in the primary visual cortex of awake ferrets ... 6
3. Properties of Inhibition Stabilized Networks (ISN) .............................................................. 8
Chapter II: Conditions to achieve surround suppression in the primary visual cortex ................. 12
1. Linear rate model with spatially invariant weight matrix ................................................... 12
2. Surround suppression constraints on the response curve .................................................... 15
3. Analytic solutions with surround suppression boundary conditions .................................. 18
4. General numerical solutions with parameter space search ................................................. 19
5. Amplification at critical filter frequency ............................................................................ 23
6. Strong surround suppression generated by stable sparse networks .................................... 25
7. Effects of different input functions with different blurring widths..................................... 27
8. Spatial oscillations in population activity ........................................................................... 29
9. Summary ............................................................................................................................. 31
~ ii ~
Chapter III: Properties of spontaneous and sensory-driven activities in the primary visual cortex
of awake ferret ........................................................................................................................... 34
1. Experimental procedure and data acquisition ..................................................................... 34
2. Principal Component Analysis of the spike trains and the dominance of a spatially long-
ranged principal component ................................................................................................. 35
3. Development of spontaneous oscillation ............................................................................ 39
4. Spontaneous oscillation in networks with surround suppression ....................................... 41
5. Modulations of the auto-covariance by sensory stimuli ..................................................... 44
6. Absence of orientation map structure in both spontaneous and sensory-driven activities . 46
7. Summary ............................................................................................................................. 50
Chapter IV: Conclusions and Discussions .................................................................................... 53
Figures in the main text................................................................................................................. 61
References ................................................................................................................................... 123
Appendix A: Structures and Functions of the Visual System..................................................... 129
1. Cortical and sub-cortical structures in the central visual pathway ................................... 129
2. Receptive field structures of neurons in the visual system. .............................................. 131
3. Columnar organization of the visual cortex ...................................................................... 134
Appendix B: Supplemental Information ..................................................................................... 136
1. Properties of Inhibition Stabilized Networks .................................................................... 136
a. Stability of the fixed point........................................................................................... 136
~ iii ~
b. Effects of increased inhibitory input ........................................................................... 138
2. Linear rate model with spatial dependency ...................................................................... 138
a. Derivation of the steady state solution ........................................................................ 138
b. Steady state solution of 2D model with circular symmetry ........................................ 140
c. Analytic solutions with surround suppression boundary conditions........................... 141
d. Expansion of the connectivity filter in the Fourier space when the network approaches
instability...................................................................................................................... 147
e. Relationship between the maximal response stimulus size and the critical stimulus size
...................................................................................................................................... 149
3. Experimental procedure and data acquisition for spontaneous and sensory-driven activity
in awake ferret V1 .............................................................................................................. 152
4. Principal Component Analysis of the activity pattern under Dark, Movie and Noise
viewing conditions ............................................................................................................. 153
a. The 1st PC mode under Movie and Noise viewing conditions ................................... 153
b. Nested model test ........................................................................................................ 155
5. Mechanisms of Hebbian amplification and properties of normal and non-normal matrices
............................................................................................................................................ 158
a. Hebbian amplification for translation-invariant linear rate models ............................ 158
b. Properties of normal and non-normal matrices ........................................................... 159
Supplemental figures .................................................................................................................. 161
~ iv ~
List of Figures
Figure 1. Typical stimulus configurations for surround suppression experiments. ...................... 62
Figure 2. Mechanisms of the Difference of Gaussian model ........................................................ 63
Figure 3. The Inhibition Stabilized Network model. .................................................................... 65
Figure 4. π subspace of numerical solutions with Gaussian input function ................................. 67
Figure 5. π subspace of numerical solutions with Gaussian input function ................................ 72
Figure 6. Histogram of the amplitude of the πΈ β πΈ connection .................................................. 75
Figure 7. Histogram of the real part of the leading eigenvalue ππΏ ................................................ 76
Figure 8. Amplification at network critical filter frequency ππΉ .................................................... 77
Figure 9. Resonance effects around the critical stimulus size ππΉ ................................................. 79
Figure 10. Simulation results of sparse networks. ........................................................................ 81
Figure 11. Effect of input blurring ................................................................................................ 84
Figure 12. Effects of Rectangular input functions ........................................................................ 89
Figure 13. Population oscillation around the critical frequency ππΉ .............................................. 95
Figure 14. Spontaneous and sensory-driven activities in the primary visual cortex of ferrets ..... 98
Figure 15. Example of Principle Component Analysis results in P129 ...................................... 100
Figure 16. Development of oscillations in spontaneous activities .............................................. 108
Figure 17. Spontaneous oscillations in the networks with surround suppression....................... 110
Figure 18. Modulations of the auto-covariance by sensory stimuli ............................................ 112
Figure 19. Oscillations in sensory-driven activities .................................................................... 116
Figure 20. Roughly power law dependency in spatial tuning curves ......................................... 118
Figure 21. Simulations from a 16-electrode array on a measured orientation map .................... 121
~ v ~
Supplemental Figure 1. PCA results of Movie viewing conditions ........................................... 162
Supplemental Figure 2. PCA results of Noise viewing conditions ............................................. 163
~ vi ~
Acknowledgements
First and foremost I want to thank my advisor Dr. Kenneth Miller, who guided and supported me
throughout my study at the Center for Theoretical Neuroscience. Ken is a great mentor with
perpetual energy and enthusiasm in research; and I would like to express my deep and sincere
gratitude to him, for his expertise, kindness, and most of all, for his patience.
The Center for Theoretical Neuroscience has been a vibrant and stimulating environment for me.
Prof. Larry Abbott and Prof. Misha Tsodyks deserve special thanks for their contributions of
time and valuable ideas. My thanks and appreciations also go to my lab buddies Dan Rubin, Xaq
Pitkow and Michael Vidne, who made my research life fun and rewarding.
I also want to thank Prof. Michael Weliky and Prof. Jozsef Fiser, who generously shared their
valuable experimental results with us. Without their help, this research would not have been
completed.
Finally, I want to thank my parents for all their love and support, and for their encouragement in
the most difficult days.
Page 1
Chapter I: Introduction and Literature Review
1. Surround suppression effects in the primary visual cortex (V1)
In the visual system, the receptive field of a neuron is the region in the visual space where a
stimulus will evoke or modify the response of that neuron. Hubel and Wiesel (Hubel and Wiesel
1959 & 1962) first explored the properties of the βclassicalβ receptive field in the primary visual
cortex, where a stimulus would evoke a direct response. The classical receptive field can be
mapped with a small optimal stimulus, typically a bar or a drifting grating. The target neuron
responds most strongly to a certain orientation of the stimulus, which is defined as the preferred
orientation of that neuron. A stimulus outside of the classical receptive field will not evoke a
direct response. Instead, the neuron's response to the center stimulus will be modulated by
stimuli in the surrounding area, which is usually referred to as the non-classical receptive field or
the extra-classical receptive field. The hierarchical organization of the visual system and the
receptive field structure are further detailed in Appendix A, Section 1 and 2.
Bar-shaped stimuli in the non-classical receptive field create length tuning effects: typically, a
bar-shaped stimulus of the target neuron's preferred orientation is placed at the center of the
receptive field; as the length of the bar increases, the response also increases and reaches its peak
at a certain optimal bar length; further increases in bar length will lead to decrease in response.
Hubel and Wiesel (Hubel and Wiesel, 1965) reported the length tuning effect with neurons in
area 18/19 (visual association areas) in cats. This length tuning effect was further examined both
in the lateral geniculate nucleus (LGN) (Levick et al., 1972) and in the primary visual cortex
(also known as V1 for βVisual Area 1β) (Dreher, 1972; Gilbert, 1977; Rose, 1977; Kato et al.,
Page 2
1978). Properties of the non-classical receptive field were studied in later experiments with more
complex stimuli. Most surround stimuli reduced the response of the target neuron, and this effect
was generally referred to as the surround suppression effect.
In a typical surround suppression experiment, the preferred orientation of the target neuron is
first determined in a preliminary search, usually with a bar-shaped stimulus. Next, a disk-shaped
drifting grating (e.g. the center stimulus in each configuration in Figure 1) is placed in the
classical receptive field of the neuron. This drifting grating is carefully tuned to have the optimal
parameters that would evoke the strongest response in the target neuron. Then the steady state
response of the target neuron is measured at different stimulus sizes. Within the classical
receptive field, the neuron's response increased with the size of the drifting grating. The response
continues to increase in the immediate surrounding area of the classical receptive field (i.e. the
βsummationβ effect), up to some optimal stimulus size. Further increases in stimulus size show a
suppressive effect and cause the response to decrease (Nelson and Frost, 1978; DeAngelis et al.,
1994; Sceniak et al., 2001; Cavanaugh et al., 2002a; Webb et al., 2005).
Typical configurations of the surround stimuli are illustrated in Figure 1: surround stimuli with
either preferred orientation or orthogonal orientation (not shown in the figure) are placed next to
the center stimulus in end-to-end, side-by-side or annulus configurations. The annulus
configuration is used to obtain an isotropic response. Flankers in end-to-end, side-by-side and
sometimes oblique configurations are used to provide a detailed map of the non-classical
receptive field (Walker et al. 1999; Cavanaugh et al. 2002b). Flanker gratings of the preferred
orientation usually induce strong surround suppression. For some cells, the strength of the
surround suppression depends on the relative location of the flanker. The suppression is strongest
in the end-to-end configuration, and becomes weaker when the flanker is presented in the side-
Page 3
by-side or oblique configuration. In general, such spatial bias is very weak, and many neurons
show the strongest suppression at an arbitrary flanker location. Flanker gratings with orthogonal
orientation (orthogonal to the preferred) induce weak or no surround suppression. Other
parameters of the configuration (e.g., spatial and temporal frequencies) also affect the strength of
the surround suppression (DeAngelis et al., 1994). In general, the strength of the surround
suppression effect is the strongest when the surround stimuli have parameters similar to that of
the optimal stimulus in the classical receptive field.
Neurons in the primary visual cortex receive feed-forward inputs from the Lateral Geniculate
Nucleus (LGN) in the thalamus. LGN neurons also have a center-surround receptive field
structure, but the characteristics of the surround suppression effects in LGN are different from
those in the primary visual cortex. Firstly, the LGN surround suppression show very weak
orientation tuning, and many neurons are not tuned for surround orientation (Kato et al., 1981;
Jones et al., 2000; Naito et al., 2007). In contrast, surround suppression in the primary visual
cortex is tuned for surround orientation, and the orientation-tuned component in LGN may also
arise from cortical feedback (Sillito et al., 2000). Secondly, the strength of LGN surround
suppression is weaker compared to that in cortex. Furthermore, neurons in upper layers of the
primary visual cortex are more likely to show strong surround suppression than neurons in the
layer that directly received LGN inputs (Jones et al., 2000; Akasaki et al., 2002). Thirdly,
experiments with dichoptic stimuli (where the center stimulus is presented to one eye and
surround stimulus to the other eye) produce significant surround suppression in the primary
visual cortex (DeAngelis et al., 1994), but such effects are very weak in LGN (Kato et al., 1981;
Xue et al., 1987). In summary, surround suppression properties in the primary visual cortex are
Page 4
most likely to emerge from cortical recurrent and feedback connections, rather than from the
feed-forward inputs.
Surround suppression effects were generally studied with a Difference of Gaussian model (DoG
model), where the inputs from the surrounding neurons to the target neuron (at the center of the
stimulus) are approximated by the difference of two Gaussian functions (Baker and Cynader
1986; Field and Tolhurst 1986; Jones and Palmer 1987a, b). Figure 2 is a schematic
demonstration of the DoG model: (top panel) the target neuron receives both excitatory (strong
and narrow, the blue Gaussian function) and inhibitory connections (wide but weak, the red
Gaussian function). The effective recurrent input to the neuron at the center is the difference of
these two Gaussian functions, shaped like a 'Mexican hat' (shown in black). The positive center
represents the classical field and the summation surround, where the response increases with
stimulus diameter. The negative surround represents the areas of surround suppression. As the
stimulus increases in size, both excitatory and inhibitory inputs to the center neuron become
stronger (bottom panel). Since the inhibitory connectivity has a wider range, the net input from
surrounding regions becomes inhibitory for large stimuli, creating the surround suppression
effect.
In the Difference of Gaussian model, the surround suppression effect arises from increases in
lateral inhibition. However, recent studies by Ozeki et al. provided strong evidence that
inhibition was actually reduced by surround stimuli in the primary visual cortex (Ozeki et al.,
2009). In their experiments, excitatory and inhibitory neuronal inputs were measured separately
in terms of excitatory and inhibitory membrane conductances. When the surround stimulus was
presented in addition to a center stimulus, there was a transient increase in inhibitory
conductance. In steady-state measurements, however, not only the excitatory but also the
Page 5
inhibitory membrane conductance was reduced by the surround stimuli. This suggests that the
target neuron receives less recurrent inhibitory inputs, despite an increase in feed-forward inputs
to the inhibitory network. Such results contradict the predictions of the DoG model, where
surround suppression arises from an increase in recurrent inhibitory inputs. This apparent
paradoxical phenomenon can be explained by an Inhibition Stabilized Network model (ISN
model). In the ISN model, the neuronal network relies on the balance between excitatory and
inhibitory populations. The excitatory sub-network is unstable by itself, but the entire network is
stabilized by the inhibitory connections (Tsodyks et al., 1997). The important features of the ISN
model are illustrated in Section 3 of this chapter.
In Chapter II, we study a linear neuronal network with Gaussian recurrent connectivity, where
the feed-forward inputs are Gaussian or Rectangular functions. We search for appropriate
parameters in the connectivity matrix so that the neurons at the center of the stimulus would
demonstrate surround suppression effects. When inhibitory connection are local (very short
ranged), the analytic results indicate that (1) the network must function in the ISN model regime;
and (2) the excitatory to inhibitory connection of the network should be longer in range than the
excitatory to excitatory connection. We also obtain numerical results from an exhaustive state
space search with biologically reasonable parameters. Most numerical solutions confirm the
analytic results, and the exceptions represent networks with only insignificant surround
suppression. In addition, many networks with strong surround suppression are characterized by a
critical frequency in the recurrent connectivity in the Fourier space. For such networks,
approximation around this critical filter frequency provides a good estimate of the peak response.
Page 6
2. Spontaneous and sensory-driven activities in the primary visual cortex of
awake ferrets
Due to its sensory nature, the activity in the primary visual cortex is believed to be
predominantly driven by the feed-forward sensory stimulus. Recently, the importance of cortical
spontaneous activity (activities in absence of stimulus) to visual information processing has
gradually received recognition. Despite its apparent randomness, spontaneous activity has
consistent spatial and temporal correlation structures (Arieli et al., 1996; Chiu and Weliky, 2001;
Kenet et al., 2003; Fiser and Weliky, 2004) that are likely to contribute to the development of
neural circuitry (Katz and Shatz, 1996; McCormick, 1999; Chiu and Weliky, 2002). For example,
cortical structures like the orientation selectivity map and horizontal connections emerge before
eye opening (Chapman et al., 1996; Durack and Katz, 1996); and maturation of such structures
can be blocked by continuous silencing of the cortex.
Spontaneous activity patterns are correlated in space and time along the visual pathway. Strong
correlations have been reported in retina and LGN (Meister et al., 1991; Wong et al., 1995;
Weliky and Katz, 1999). In cortex, spontaneous activity patterns show strong correlation with
maps evoked by sensory stimuli (Tsodyks et al., 1999; Kenet et al., 2003; Fiser et al., 2004). This
correlation can be modeled as selective amplification of the activity pattern in the neuronal
network, evoked by an oriented stimulus. Such networks are typically constructed under Hebbβs
rule, where neurons with similar firing patterns have a tendency to excite one another, while
opposite firing patterns lead to a mutually inhibitory connection (Douglas et al., 1995; Seung,
2003; Goldberg et al., 2004). In such networks of strong recurrent connectivity, inputs to certain
selected patterns can be selectively amplified by having a much slower decay rate. Unamplified
patterns decay at a much faster rate, determined by the synaptic time constant of the neuron. If
Page 7
the inputs have no bias toward any pattern and the network is stable so that with time no pattern
grows to infinity, then the pattern with the slowest decay rate will emerge from the spontaneous
activity by accumulating to the highest amplitude, i.e. the Hebbian amplification effect.
In Chapter III, we study recordings in the primary visual cortex of awake and free viewing
ferrets at different age groups (Chiu and Weliky, 2001). The experiments measure both
spontaneous activities (under complete darkness) and sensory-driven activities (by correlated
inputs from a natural scene movie and uncorrelated noise). The differences between sensory-
driven activities and spontaneous activities decrease as the animal matured, in agreement with
previous studies (Chiu and Weliky, 2001; Fiser and Weliky, 2004). In young animals, correlated
inputs significantly increase the temporal correlation in the auto-correlation of the activity
patterns. In the mature age group, however, such modulations by sensory inputs become much
smaller. We apply Principal Component Analysis (PCA) over the normalized data. The first
Principal Component (PC), which contributes the most to the total variance of the activity pattern,
is a spatially homogeneous (βDCβ) and temporally slowly-varying mode. For mature animals,
this mode is the only slowly-decaying mode, with all other modes quickly decaying to the
background level.
Recordings from the mature groups show that oscillations of 8~14Hz emerge from the auto-
correlation structures in both the spontaneous and sensory-driven activities. The noise stimuli
induce a strong and persistent oscillation around 10 Hz in the late age group of postnatal
129~168 days. Neuronal oscillations in human and animal studies are behavior dependent, and
serve important computational functions in perception, memory and cognition (Ward, 2003;
Cooper et al., 2003; Buzsaki and Draguhn, 2004). In general, oscillations of 8~14Hz fall in the
-band of brain waves. -band waves usually represent activities of the visual cortex in an idle
Page 8
state. Studies by Kelly et al. (Kelly et al., 2006) showed that oscillation, especially
synchronization, could be attributed to the suppression of competing distractions. An -like
variant in motor cortex, i.e. the rhythm (8~13 Hz), has been argued to represent an βidleβ or
βdisengagedβ state (Fontanini and Katz; 2005).
With proper simplifications, the spontaneous activities can be studied by a linear rate model with
both excitatory and inhibitory populations, as in the surround suppression study. The
spontaneous activities have relatively lower firing rates compared to the sensory-driven activities,
thus we model the spontaneous activity as perturbations around the network fixed point. As
mentioned in the previous section, neuronal circuitries in the primary visual cortex also
demonstrate surround suppression effects. Therefore, in the model, the possible parameters of the
recurrent weight matrices are given by the numerical results of parameters showing surround
suppression in Chapter II. In many of such networks, a spatially DC component with temporal
oscillation and large decay time constant emerges as a result of Hebbian amplification. This
spatially DC component closely resembles the 1st PC mode in the experiment. In addition, the
characteristics of the 1st PC mode suggest that such networks function in the ISN regime.
3. Properties of Inhibition Stabilized Networks (ISN)
In this paper, we study neural networks with both excitatory and inhibitory populations. In
general, the change in firing rate of the neurons depends on the external and recurrent input plus
a self-decay term; thus a general model of firing rate can be constructed as in Equation 1:
ππ
π
ππ‘πΈ = βπΈ + ππ(ππππΈ β πππ πΌ + ππ)
Page 9
ππ
π
ππ‘πΌ = βπΌ + ππ(ππππΈ β ππππΌ + ππ)
(Eq. 1)
where πΈ and πΌ are the averaged firing rates of the excitatory and inhibitory population. ππ and ππ
are the neuronal response functions that modulate the effective inputs received by the neurons. ππ
and ππ are the membrane time constants; ππ and ππ are external inputs to the excitatory and
inhibitory populations. ππ₯π¦ in the weight matrix represents the connection from population y to
population x.
The response function usually takes the form of a rectified linear function or a sigmoid function.
We used the generalized logistic functions for this illustration (shown in Figure 3a):
π π₯ = πΎ
1 + ππβπ΅(π₯π
βπ) 1/π£
where πΎ = 1.0, π = 0.5, π΅ = 1.5, π = 7.0, π = 3.0 and π£ = 0.5. Figure 3b-e are the phase
planes of Equation 1, where the excitatory and the inhibitory nullclines (given by π
ππ‘πΈ = 0 and
π
ππ‘πΌ = 0) are shown in blue and red respectively. Points on the excitatory nullcline represent
fixed points of the excitatory sub-network when the inhibitory firing rates are clamped at given
values; while points on the inhibitory nullcline represent fixed points of the inhibitory sub-
network with clamped excitatory firing rates. The stability of each nullcline β meaning whether
the fixed points of a sub-network are stable when the firing rate of the other sub-network is
clamped β depends on the slope of the nullcline, as indicated by blue and red arrows in Figure 3b
and 3c (see Appendix B Section1a for mathematical details). The inhibitory nullcline always has
Page 10
a positive slope and is always stable. The excitatory nullcline can have either positive or negative
slope: portions with negative slope are stable; those with positive slope are unstable. The fixed
point of the network is located at the intersection of the two nullclines. For the fixed point to be
stable, the slope of the excitatory nullcline must be smaller than that of the inhibitory nullcline.
Given that the fixed point is stable, there are two different scenarios depending on the slope of
the excitatory nullcline at the fixed point, as illustrated in Figure 3: Inhibition Stabilized
Networks (ISN, Figure 3c, 3e) vs. non-ISN (Figure 3b, 3d). In the non-ISN scenario, the
excitatory to excitatory connection is weak (Figure 3b, network parameters: πππ = 0.15,
πππ = 0.7, πππ = 2.0, πππ = 1.0, ππ = 0.7, ππ = 0.0). The excitatory nullcline has negative
slope and the inhibitory nullcline has positive slope. Both excitatory and inhibitory nullclines are
stable, and the fixed point of the network is also stable. In the ISN scenario, the excitatory to
excitatory connection is much stronger (Figure 3c, network parameters: πππ = 0.75, πππ = 0.4,
πππ = 1.7, πππ = 0.75, ππ = 0.3, ππ = 0.0). The excitatory nullcline has a segment of positive
slope (shown as the broken line), and is unstable by itself. However, with recurrent inhibition,
the fixed point of the network remains stable, i.e. the network is stabilized by inhibition. The
stability of the examples in Figure 3 is checked by calculating the eigenvalues of the linearized
equation around the fixed point.
One major difference between the ISN and non-ISN scenarios is in the shift of the network fixed
point with increased external input to the inhibitory population. In both scenarios, increased
input to the inhibitory population shifts the inhibitory nullcline upward. For the non-ISN, the
new fixed point has decreased firing rate of the excitatory population and increased firing rate of
the inhibitory population (Figure 3d, network parameters: ππ increased from 0.0 to 0.15, other
parameters same as in Figure 3b). In the ISN scenario, however, the excitatory nullcline has a
Page 11
positive slope. Despite the additional inputs to the inhibitory population, the new fixed point has
lower firing rates for both excitatory and inhibitory populations (Figure 3e, network parameters:
ππ increased from 0.0 to 0.15, other parameters are the same as in Figure 3c). This additional
input inhibits the excitatory activity, and the network shifts to a less active state due to
withdrawal of excitation.
Wilson and Cowan (Wilson and Cowan, 1972) studied a recurrent network model with excitatory
and inhibitory populations, and illustrated that a stable fixed point can exist on the unstable
branch of the excitatory nullcline. With stronger recurrent excitatory connectivity, this fixed
point will become unstable, and limit cycles will appear in the phase plane. The limit cycles have
been proposed as the underlining mechanism of network oscillations in hippocampus (Leung,
1982; Tsodyks et al., 1996). Tsodyks et al. (Tsodyks et al., 1997) demonstrated the effect that
external input to inhibitory neurons reduced their responses due to the withdrawal of network
excitation. This paradoxical effect of the ISN has been observed in recent surround suppression
experiments with direct membrane conductance measurements in the primary visual cortex
(Ozeki et al., 2009). In our work, we use a linearized version of the rate model in Equation 1 to
study networks where the connectivity weights between neurons depend only on their relative
locations. We apply this model to study surround suppression effects and search for networks
with appropriate parameters. The same model is also applied in the study of spontaneous
activities in awake ferrets. Both studies lead to network solutions in the ISN regime, suggesting
that the ISN mechanism might play an important role in the neural circuitry in the primary visual
cortex.
Page 12
Chapter II: Conditions to achieve surround suppression in
the primary visual cortex
1. Linear rate model with spatially invariant weight matrix
In this chapter, we study a network of excitatory and inhibitory complex cells in the primary
visual cortex with the same preferred orientation. The stimuli are optimized drifting gratings of
variable sizes at the preferred orientation. Complex cells in the primary visual cortex respond to
the contrast of the stimuli; therefore a steady drifting grating stimulus of a given size effectively
provides a constant feed-forward input. We apply a linearized rate model with spatial
dependency in the weight matrix, similar to the generic rate model given by Equation 1:
ππ
ππ‘ πΈ(π₯ )
πΌ(π₯ ) = β
πΈ(π₯ )
πΌ(π₯ ) + π π₯ β²π( π₯ β² β π₯ )
πΈ(π₯ β²)
πΌ(π₯ β²) +
ππ(π₯ )
ππ(π₯ )
(Eq. 2)
where πΈ π₯ and πΌ π₯ are firing rates of the excitatory and inhibitory neurons respectively.
π = ππ 00 ππ
is the time constant matrix with excitatory and inhibitory membrane time
constants ππ and ππ .
Since all neurons in the network share the same preferred orientation, the connectivity weight
between two neurons depends only on their relative positions. For complex cells, the spatial
connectivity weights can be approximated by weighted Gaussian functions on their receptive
Page 13
fields (Stepanyants et. al. 2009). For simplicity, we assume the connectivity weights are given by
Gaussian functions:
π( π₯ β² β π₯ ) =
πππππ₯π(β π₯ β² β π₯ 2
2πππ2
) βπππππ₯π(β π₯ β² β π₯ 2
2πππ2 )
πππππ₯π(β π₯ β² β π₯ 2
2πππ2 ) βπππππ₯π(β
π₯ β² β π₯ 2
2πππ2 )
(Eq. 3)
The sub-index of ππ₯π¦ denotes connection from the y population to the x population. ππ(π₯ )
ππ(π₯ ) is
the effective external input, centered at the origin π₯ = 0. The drifting grating stimulus provides a
steady input to the complex cells, thus the input term is independent of time. In this study, we
use two sets of input functions (Gaussian and Rectangular input functions) to simulate stimuli
with different edge conditions. In addition, we assume this feed-forward input is blurred along
the visual pathway, by a Gaussian function of width π0. Thus the effective input functions to the
primary visual cortex are:
Gaussian input: ππ(π₯ )
ππ(π₯ ) =
π
π2+π02ππ₯π β
π₯ 2
2(π2+π02)
ππ
ππ
(Eq. 4.1)
Rectangular input: ππ(π₯ )
ππ(π₯ ) =
1
2 πππ(
π₯ +π
2π0) β πππ(
π₯ βπ
2π0)
ππ
ππ
(Eq. 4.2)
Page 14
where π is the size of the input, corresponding to the half width of the length tuning or the radius
of the circular stimulus; π0 comes from the Gaussian blurring. ππ and ππ are the relative
intensities of the inputs to the excitatory and the inhibitory neurons. πππ(π₯) is the Gaussian error
function. We use the Gaussian input function to obtain analytic results. Both types of input
functions are used for numerical simulations.
For length tuning experiments and experiments with lateral flanker setup, the network can be
modeled by a 1-dimensional spatial array along the lateral direction. For the circular stimulus,
the network can also be reduced to a 1-dimensional system of radial location from the center of
the stimulus. Both cases can be generalized by the same 1-dimensional model (the details of the
derivation can be found in Appendix B, Section 2a). We solve for the steady state solution of
Equation 2. The firing rates of the neurons at the center of the stimulus (π₯ = 0) are determined
by the inverse of the connectivity matrix in the Fourier space (See Appendix B, Section 2b for
details), i.e.:
πΈ(0)πΌ(0)
=1
2π ππ πΌ β π (π)
β1 π π(π)
π π(π)
(Eq. 5)
where πΌ is the identity matrix, π (π) and π π(π)
π π(π) are the weight matrix and input functions in
the Fourier space.
The weight functions of the networks are translation-invariant as shown in Equation 3, and the 4
sub-matrices of ππ₯π¦( π₯ β² β π₯ ) are diagonalized simultaneously by the Fourier bases. As a result,
for a given spatial frequency π, the network dynamic is determined by a 2-by-2 matrix:
Page 15
π (π) =
π ππππ₯π(β
πππ2 π2
2) βπ ππππ₯π(β
πππ2 π2
2)
π ππππ₯π(βπππ
2 π2
2) βπ ππππ₯π(β
πππ2π2
2)
(Eq. 6)
Here, π π₯π¦ are amplitudes in the Fourier space. The upper-left πΈ β πΈ term captures the recurrent
connectivity within the excitatory sub-network. We classify the network to be an inhibition
stabilized network, if (1) it is globally stable, and (2) for some spatial frequency the excitatory
sub-network is unstable, i.e.:
π ππππ₯π βπππ
2 π2
2 > 1
for some k, which is true if and only if π ππ > 1.
2. Surround suppression constraints on the response curve
The firing rates of the center neurons in Equation 5 depend on stimulus size π. In this study, we
aim at finding biologically reasonable parameters so that the response curve of the target neuron
in the model would have the same characteristics as in surround suppression experiments.
Typical response curves in such experiments comprise a summation area for small input sizes, up
to some summation peak, and a suppression area beyond that point. The size of the summation
area is usually larger than or comparable to the size of the classical receptive field. In general,
the suppressive surround of the response curve may take various shapes; however, the
asymptotic response for stimulus size infinity should be weaker than that of the summation peak.
Furthermore, even with suppressive inputs from the surround areas, the center neuron still
Page 16
receives considerable amount of feed-forward inputs; and the response to stimuli covering the
suppressive surround is still stronger than baseline activity with no stimulus at all (the baseline
activity is zero in our linear model).
Thus we generalize the surround suppression constraints on the response curve as follow:
A. Suppressive surround: we require that the response curves πΈ(π₯ = 0, π)πΌ(π₯ = 0, π)
of the neurons at
the center of the stimulus to show a global peak at ππ , i.e. the maximal response stimulus size,
and 0 < ππ < β. We generally consider stimulus size π < ππ to be in the summation region of
the response curve, and stimulus size π > ππ to be in the suppressive surround, regardless of the
specific shape of the response curve.
B. Non-negative response: surround suppression reduces the firing rate but does not drive the
network below its baseline activity (the spontaneous firing rate). Thus we require that
πΈ(π₯ = 0, π)πΌ(π₯ = 0, π)
> 0 for all stimulus sizes π.
C. Stability: the network dynamics given by Equation 2 must be stable. Therefore, in the Fourier
space, all eigenvalues of the connectivity matrix π (π) β πΌ must have negative real part for all
k. For any 2-by-2 matrix at a given π (Equation 6), this means the determinant of the
connectivity matrix is positive while the trace is negative.
In many experiments, the response in the summation center increases monotonically with the
stimulus size; and for stimuli large enough, the response monotonically decreases with the
stimulus size (DeAngelis et al., 1994; Sengpiel et al., 1997; Akasaki et al., 2002; Angelucci et al.,
2002; Ozeki et al., 2004). In the numerical studies (Section 4 to 8 of this chapter), we search for
Page 17
general cases without the constraint of monotonically decreasing response curve, i.e. solutions
with oscillations in the response curve are allowed. For simplicity, we assume this additional
constraint in the analytic studies (Section 3). Thus, Constraint A can be reduced to analytic
boundary conditions:
A'. Summation for small stimulus: the response curve increases with the stimulus size when
the stimulus size is small:
π
ππ πΈ(0, π β 0)πΌ(0, π β 0)
> 0
A''. Monotonic suppression for large stimulus: the response curve decreases for stimuli large
enough:
π
ππ πΈ(0, π β β)πΌ(0, π β β)
< 0
Under these conditions, the response curve has a single global peak stimulus size ππ . For
stimulus smaller than ππ , the response is always non-negative. For stimulus larger than ππ , as
long as the asymptotic response at stimulus size infinity is non-negative, all responses in the
suppressive surround stay above the activity baseline. This leads to a simplified boundary
condition for non-negative response:
B'. Non-negative response for large stimulus:
πΈ(0, π β β. )πΌ(0, π β β. )
> 0
Page 18
3. Analytic solutions with surround suppression boundary conditions
With Gaussian input functions, we are able to obtain analytic solutions for the model in Section 1,
with the boundary conditions listed in the previous section. The detailed derivations are shown in
Appendix B Section 2c. The results are conditions on the model parameters as follow, equivalent
to the constraints in the previous section:
Solutions under Constraint A':
ππ πΌ β π π β1
β πΌ πΆππΆπ
β
ββ
> 0
Solutions under Constraint A'':
1 + π ππ βπ ππ
π ππ 1 β π ππ
π πππππ
2 βπ πππππ2
π πππππ2 βπ πππππ
2 1 + π ππ βπ ππ
π ππ 1 β π ππ
πΆππΆπ
< 0
Solutions under Constraint B':
1 + π ππ βπ ππ
π ππ 1 β π ππ
πΆππΆπ
> 0
Solutions under Constraint C:
π·ππ‘ π π β πΌ > 0 and ππ π π β πΌ < 0, for every π.
Combining the π β β results in A'' and B', we have:
π
ππ πΈ 0, π β β
πΌ 0, π β β =
1 + π ππ βπ ππ
π ππ 1 β π ππ
π πππππ
2 βπ πππππ2
π πππππ2 βπ πππππ
2 πΈ 0, π β β
πΌ 0, π β β < 0
Page 19
For the inhibitory neuron at the center of the stimulus:
πππ2 β
π ππ β 1
π ππ
πππ2 πΈ 0, π β β β
π ππ
π ππ
πππ2 β
π ππ β 1 π ππ
π πππ ππ
πππ2 πΌ 0, π β β < 0
(Eq. 7)
In general, the inhibitory connections are shorter in range compared to the excitatory connections
(Gilbert and Wiesel, 1983; Das and Gilbert, 1995). Excitatory neurons can form long-range
lateral connections (Kisvarday et al., 1997; Azouz et al., 1997, Sceniak et al., 2001, Stettler et al.,
2002) while inhibitory neurons lack such long-range connections. In addition, Equation 7
depends on the square of connection widths. When the inhibitory connections are local, the
contribution of inhibitory term becomes insignificant compared to that of the excitatory term.
Therefore, we immediately have π ππ > 1, and the network functions in the ISN regime. In
addition, we also have:
πππ
πππ
2
>π ππ
π ππ β 1> 1
Thus πππ > πππ , which means the πΌ β πΈ connections are longer in range than the πΈ β πΈ
connections.
4. General numerical solutions with parameter space search
In this section we obtain numerical solutions given by the Constraints A, B and C in Section 2
without simplifications or approximations for the analytic solutions. The width of πΈ β πΈ
connection is set to be the unit length πππ β‘ 1; and there are 9 free parameters in the model: the
Page 20
widths of the connections: πππ , πππ , πππ ; the amplitudes of the connections: πππ , πππ , πππ , πππ ; the
width of the input blurring π0; and the ratio of the intensities of the input to the excitatory and the
inhibitory neurons: ππ/ππ. We perform an exhaustive parameter space search and check the
resulting response curves against the surround suppression constraints. The parameter space is
chosen to have a wide range, covering biologically reasonable cases, but we do not restrict the
parameter space specifically to contain only biologically reasonable solutions. The πΌ β πΈ
connection width is chosen to be comparable to πΈ β πΈ connection width, up to about twice as
wide: πππ = (0.7, 0.9, 1.1, 1.3, 1.5, 1.7, 1.9). In general, inhibitory connections are shorter ranged
than the excitatory connections: πππ , πππ = (0.05, 0.1, 0.3, 0.5, 0.7, 0.9, 1.3), we include πππ , πππ =
0.05 to account for local inhibitory connection. Although biologically unlikely, we also allow
πππ , πππ = 1.3 so that we do not exclude the possibility of long range inhibitory connections. The
amplitudes of excitatory connections are given by πππ , πππ = (0.20, 0.35, 0.50, 0.65, 0.80).
Since πππ = 1 and π ππ = 2πππππππ , both non-ISN (π ππ < 1) and ISN (π ππ > 1) solutions are
possible in the simulation. The amplitudes of inhibitory connections have a wide range:
πππ , πππ = (0.1, 0.2, 0.4, 0.8, 1.6, 3.2, 6.4), so that in the local inhibitory connection case, the
area under the envelope of the inhibitory Gaussian weight functions (given by ππππππ and ππππππ )
could still be comparable to that of the excitatory ones. The width of the blurring π0 ranges up to
the width of the πΈ β πΈ connection: π0 = (0.0, 0.25, 0.5, 1.0). The ratio of the inputs to
excitatory vs. inhibitory neurons is: ππ/ππ = (0.5, 1.0, 2.0). For each parameter combination, the
stability of the network is checked first. Then for the stable networks, the response curve is
computed numerically and checked against Constraint A and B. Numerical results are obtained
for both Gaussian and Rectangular input functions.
Page 21
We first compare the numerical results to the analytic results with local inhibition in the previous
section. We set the inhibition width to be short ranged: πππ , πππ = 0.05. Since the previous
analytic result suggested long range πΌ β πΈ connection (πππ > πππ ), the range of πππ is extended
to up to about 4.0 for better illustration. The ranges of the other parameters remain unchanged.
Figure 4a is an example with Gaussian input function, ππ/ππ = 1.0 and no blurring π0 = 0. The
numerical results agree with analytic ones: all solutions are in the ISN regime (π ππ > 1) with
long range πΌ β πΈ connection (πππ > πππ ).
Next, we examine the general numerical solutions over the entire parameter space. To simplify
the illustration, we first plot the results in the density maps of the connection widths πππ , πππ and
πππ in the 3-dimensional π subspace. Figure 4b is an example with Gaussian input function,
ππ/ππ = 1.0 and π0 = 0, color coded by solution density. The solution density increases as πππ and
πππ become larger.
For each solution in the parameter space, we quantify the strength of surround suppression by
calculating the Suppression Index (ππΌ), defined as the ratio of the peak-to-infinity suppression vs.
the peak response:
ππΌ =π ππ ππππ πππ
β π ππ ππππ ππββ
π ππ ππππ πππ
where ππ is the stimulus size at maximal response. The suppression indices are calculated for the
excitatory and the inhibitory populations respectively (ππΌπ , ππΌπ). When used without sub-indices,
ππΌ is the minimum of ππΌπ and ππΌπ . We consider solutions with ππΌ < 5% to be insignificantly
surround-suppressed and solutions with ππΌ β₯ 50% to be strongly surround-suppressed. Figure 4c
and 4d are density maps of the 3-dimensional π subspace, color coded by the averaged ππΌ of the
Page 22
excitatory and the inhibitory responses respectively. In general, excitatory response curves show
stronger surround suppression compared to the inhibitory ones (different levels in color coding
legend), and solutions with strong surround suppression tend to cluster in the region with large
πππ and πππ .
In Figure 4b-4d, the solution density only weakly depends on the width of the πΌ β πΌ connection
πππ . Thus we project the 3-D map on the πππ vs. πππ plane, as shown in Figure 4e. Excitatory
neurons can effectively inhibit neighboring excitatory neurons via an πΈ β πΌ β πΈ connection
chain, where each excitatory neuron projects to its neighboring inhibitory neurons and the
inhibitory neurons further suppress their neighboring excitatory neurons. The width of such
lateral inhibition is πππ2 + πππ
2 :
ππ₯β² exp β π₯ β π₯ β² 2
2πππ2 exp β
π₯β²2
2πππ2 β exp β
π₯2
2 πππ2 + πππ
2
(Eq. 8)
The solid curve in Figure 4e represents πππ2 + πππ
2 = πππ2 = 1; all solutions are above this curve,
i.e. the lateral inhibition has longer range compared to the lateral excitation by πΈ β πΈ connection.
The region above the broken line is given by the biological constraint πππ > πππ . Within this
region, all solutions satisfy πππ > πππ , i.e. the πΌ β πΈ projection is wider than the πΈ β πΈ
projection. In general, surround suppression solutions favor large πππ ; separate search results
with wider range of πππ show that surround suppression solutions are abundant when πππ > 2πππ .
Next, we examine the connections amplitudes in the Fourier space: π ππ , π ππ , π ππ and π ππ . To
illustrate the results in a 3-dimentional map, we normalize all amplitudes by the amplitude of the
Page 23
πΈ β πΈ connection. Figure 5a is the 3-dimensional map of the relative amplitudes in Fourier
space: π ππ /π ππ , π ππ /π ππ and π ππ/π ππ , color coded by solution density. Figure 5b and 5c are the
same map color coded by suppression index of excitatory and inhibitory responses respectively.
The plane in Figure 5a is a least squares fit of the data, given by π ππ = 2.10π ππ + 0.35π ππ β
1.21π ππ .
The amplitude of the πΈ β πΈ connection π ππ determines whether the network is an ISN or not.
Figure 6 is the histogram of π ππ , grouped and color coded by the minimal suppression index ππΌ.
Most solutions are in the ISN regime, more than 88% of all solutions have π ππ > 1, and most of
the non-ISN solutions only show insignificant surround suppression (ππΌ < 5%). Apart from such
solutions, only less than 1% of the total solutions are non-ISN. Stronger surround suppression
favors large amplitude of the πΈ β πΈ connection. For groups with ππΌ β₯ 10%, most solutions have
very strong recurrent connections in the excitatory sub-network, i.e. π ππ β₯ 1.625.
5. Amplification at critical filter frequency
The stability of the weight matrix π is checked for all spatial frequencies. The stability of π (π)
at a given spatial frequency π is determined by the largest real parts of all eigenvalues of π (π).
We define the eigenvalue with the largest real part across all spatial frequencies to be the 'leading
eigenvalue'. Figure 7 is the histogram of the real part of this leading eigenvalue of π (π) for all
numerical solutions, color coded by the suppression index. Networks with strong surround
suppression (ππΌ β₯ 50%) are generally closer to instability, as the real part of the leading
eigenvalue is close to 1. We define the critical filter frequency ππΉ to be the frequency
Page 24
corresponding to the leading eigenvalue of the inverse of the connectivity matrix πΌ β π (π) β1
.
From a signal processing point of view, the neuronal response in Equation 5 is the result of the
input signal π π(π)
π π(π) passing through the band-pass filter πΌ β π (π)
β1. When the network
approaches instability, π·ππ‘ πΌ β π ππΉ β 0, and the inverse of the connectivity matrix
πΌ β π (π) β1
has a sharp peak around ππΉ , providing a large amplification for inputs with
frequencies comparable to ππΉ .
We expand the filter πΌ β π (π) β1
as a series in π·ππ‘ πΌ β π π ; dropping the higher order
terms (See Appendix B Section 2d for details), we arrive at:
πΌ β π π β1
~1
π·ππ‘ πΌ β π π
1 + π ππ π
π ππ π 1 + π ππ π βπ ππ π
(Eq. 9)
In Equation 9, the filter acts as a projection operator, mapping the input vector 1 + π ππ π
βπ ππ π
into the output vector 1 + π ππ π
π ππ π . For Gaussian input functions, the input at frequency ππΉ is
π π(π, ππΉ)
π π(π, ππΉ) = 2πππβππΉ
2 π2+π02
2 ππ
ππ , which reaches its maximum at the critical stimulus
size ππΉ β‘ 1/ππΉ, as illustrated in Figure 8 top panel. As long as the stimulus is not orthogonal to
the input vector, i.e. 1 + π ππ(π) βπ ππ (π) ππ
ππ β 0, the response curve would reach its peak
around the critical stimulus size, as illustrated in Figure 8 bottom panel: ππ ~ππΉ, where ππ is the
maximal response stimulus size. In addition, for solutions with strong surround suppression
Page 25
(ππΌ β₯ 50%), the response at critical stimulus size πΈ 0, ππΉ
πΌ 0, ππΉ should be linearly dependent on
the filter output vector 1 + π ππ ππΉ
π ππ ππΉ around the critical filter frequency ππΉ . Figure 9a shows
the ratio of the responses at the critical stimulus size πΈ 0, ππΉ /πΌ 0, ππΉ vs. the ratio of the
output vector 1 + π ππ ππΉ /π ππ ππΉ for all solutions with ππΌ β₯ 50% (286 solutions). The
result shows a linear dependency: the solid line is the linear least squares fit, given by πΈ/πΌ =
0.87 1 + π ππ /π ππ β 0.08, π2 = 0.94.
A saddle point approximation around the critical filter frequency ππΉ show that, for solutions with
weaker surround suppression, the maximal response stimulus size ππ is generally larger than the
critical stimulus size ππΉ (the mathematical details are given in Appendix B, Section 2e). Figure
9b is the scatter plot of the maximal response stimulus size ππ vs. the critical stimulus size ππΉ for
all solutions. In general, most solutions satisfy ππ β₯ ππΉ; and for solutions with strong surround
suppression (ππΌ β₯ 50%.), ππ is very close to ππΉ .
6. Strong surround suppression generated by stable sparse networks
In the above section, networks with strong surround suppression tend to approach instability. To
improve stability, we introduce random sparseness to the connectivity matrix; and across several
random instantiations, we choose the ones whose real part of the leading eigenvalue is smaller
than that of the original dense matrix. Despite this reduction in eigenvalue, some neurons in
these sparse networks show very strong surround suppression effects.
Page 26
Sparseness in the connections is modeled in the weight matrix as a form of variability, while the
overall envelope of connectivity functions remains unchanged. We simulate networks of 4000
neurons (2000 excitatory and 2000 inhibitory) with densely connected matrices given by
different parameters in Section 4. We randomly set the weights in the 'dense' matrix to 0 with
probability 1 β π , where π is the sparseness factor (π = 5%, 10%, 20%). The resulting sparse
matrix is normalized so that the sum of the excitatory weights and the sum of the inhibitory
weights to each cell remain the same as the dense matrix. In the previous sections, the weight
matrix was translation-invariant; therefore all neurons had the same response curve and ππΌ value.
Sparseness breaks the translation-invariance, and neurons in the sparse network have different
responses curves and ππΌ values.
Figure 10a shows the population distribution of the ππΌ values in two different scenarios where
the population ππΌ values of the dense matrices are at different levels. Population 1 corresponds to
a case of low population ππΌ in the dense matrix, where the averaged population suppression
indices are ππΌπ = 33% and ππΌπ = 21%, for the excitatory and the inhibitory sub-populations. The
simulation parameters are: πππ = 0.5, πππ = 1.9, πππ = 0.3, πππ = 0.65, πππ = 0.4, πππ = 0.5,
πππ = 0.4, ππ/ππ = 1.0, π0 = 0 and π = 20%. The distribution is unimodal, similar to the
experimental results reported by Walker et al (Walker et al., 2000). Population 2 is another
sparse network with stronger surround suppression in the dense matrix (ππΌπ = 58% and ππΌπ =
48%), simulation parameters: πππ = 0.5, πππ = 1.9, πππ = 0.3, πππ = 0.8, πππ = 0.8, πππ =
0.5, πππ = 0.4, ππ/ππ = 1.0, π0 = 0 and π = 20%. The distribution is bimodal, similar to the
experimental results with a heavy tail at high ππΌ values (Jones et al., 2000; Akasaki et al., 2002).
Neither sparse matrix is close to instability, since the real parts of their leading eigenvalues
satisfy ππππ ππΏ < 0.9. Nonetheless, both examples contain neurons with very strong surround
Page 27
suppression (ππΌ β₯ 90%). In contrast, when the same sparseness is introduced to dense networks
with insignificant ππΌ values (ππΌ = 2%, for population 3 and population 4), all neurons in the
sparse network have low ππΌ values even when the network approaches instability, i.e. the real
part of the leading eigenvalues is very close to 1: ππππ ππΏ > 0.95 (Figure 10b).
7. Effects of different input functions with different blurring widths
The previous sections focused on the results with Gaussian input function without blurring
(π0 = 0). In this section, we examine surround suppression effects under different stimulus
conditions: two types of input functions (Gaussian or Rectangular) with three different blurring
widths: π0 = (0.25, 0.5, 1.0).
Figure 11 shows an example of the results with input blurring, where the input function is
Gaussian and the blurring width π0 = 0.25. The input blurring increases the number of solutions
(22268 solutions in the blurring result vs. 7199 in the non-blurring result). Figure 11a, 11b and
11c show the input blurring solutions in the π subspace and π subspace. These new solutions
have a similar structure in the π subspace compared to the solutions without input blurring. In
the π subspace, the solutions cover a larger range along each axis. The red markers in Figure
11c represent solutions that also appear in the results without blurring. Such solutions tend to
cluster in the local inhibition region where both πΌ β πΈ and πΌ β πΌ connections are short in range.
In results with no input blurring, solutions with strong surround suppression were characterized
by the critical filter frequency ππΉ , and ππ ~ππΉ, as shown in Figure 9b. In Figure 11d, solutions
with input blurring are plotted in the same way, color coded by the suppression index. Here,
solutions with strong surround suppression (dark red) can be roughly divided into two clusters.
Page 28
The first cluster along the diagonal of ππ = ππΉ corresponds to the solutions obtained without
input blurring. The second cluster contains solutions whose maximal response stimulus size ππ is
small and is independent of the critical stimulus size ππΉ .
The second cluster of solutions represents networks where the recurrent input to the central
neuron is globally inhibitory. Such networks generate monotonically decreasing response curves
in the absence of input blurring. With input blurring, the feed-forward input received by the
center neuron is the stimulus convolved with the blurring function. When the stimulus size is
much smaller than the width of the input blurring, the convolution is roughly proportional to the
sum of the stimulus over positions. Thus, the response curves of the center neurons always show
an initial summation. When the stimulus size becomes larger, this feed-forward summation effect
becomes weaker than the recurrent inhibition effect; and the response becomes surround-
suppressed. In these networks, the initial summation is short in range compared to the width of
πΈ β πΈ connection (πππ = 1).
Figure 12 shows the results with Rectangular functions without input blurring. There are more
solutions (84487 solutions) than that of the Gaussian input functions (7199 solutions). These
solutions span the entire π subspace, as shown in Figure 12a and 12b. Most of the new solutions
have very small πππ (Figure 12c top panel). Such solutions generate monotonically increasing
response curves with Gaussian input functions, i.e. the maximum of the response curve is at
π β β . At π β β the population response depends on the inverse Fourier transform of the
input function (Appendix B, Section 2c), hence the response of the center neuron (π₯ = 0) is
given by the integral π (π)β
ββππ. The Fourier transform of a Rectangular function is the Sinc
Page 29
function. The area under the central peak of the Sinc function ([β1/π, 1/π]) is larger than the
entire integral over the real line:
π πππππ
πππ
1/π
β1/π
ππ~1.18 π πππππ
πππ
β
ββ
ππ
Most of these solutions do not show resonance in the connectivity filter as describe in Section 5,
and their connectivity filters peak at π = 0. Thus, a global peak in the response curve may
appear when the central peak of the Sinc function is optimal for the connectivity filter, and such
solutions become surround suppression solutions as defined in Constraint A, Section 2.
Most of these new solutions have very insignificant surround suppression (ππΌ < 1%); and the
solutions with significant surround suppression (ππΌ β₯ 5%) satisfy πππ2 + πππ
2 > πππ2 = 1 (Figure
12c bottom panel), similar to the results with Gaussian input functions. Such similarities in
distribution are also seen in the π subspace (Figure 12d and 12e), for solutions with large ππΌ.
8. Spatial oscillations in population activity
As mentioned in Section 5, many surround suppression solutions (especially the ones with very
strong ππΌ) have a sharp peak in the filter πΌ β π (π) β1
around the critical frequency ππΉ . Figure
13a shows the histograms of the critical frequency ππΉ for solutions with Gaussian input and no
blurring. The top panel contains all 7199 solutions with ππΌ > 0% and most solutions have
ππΉ > 0 (7161 out of 7199). The bottom panel shows solutions with ππΌ β₯ 50% (286 solutions),
and all solutions have ππΉ > 0. The population activity is given by the inverse Fourier transform
of the filtered input in the Fourier space (Equation S7, Appendix B, Section 2a). Therefore, the
Page 30
sharp peak at ππΉ > 0 in the filter πΌ β π (π) β1
can lead to spatial oscillations in the population
activity. Figure 13b illustrates population activity patterns of an example network at various
stimulus sizes (network parameters: πππ = 0.5, πππ = 1.9, πππ = 0.3, πππ = 0.65, πππ = 0.4,
πππ = 0.5, πππ = 0.4, same as Population 1 in Section 6), and the critical stimulus size is
ππΉ = 1.1. When the stimulus size is small, the population response is localized around π₯ = 0. As
the stimulus size increases, a population oscillation emerges and becomes strong when the
stimulus size is comparable to the critical stimulus size. When the stimulus size becomes very
large, the population response becomes constant and the oscillation disappears.
For Gaussian input functions, there is one global critical stimulus size ππΉ corresponding to the
resonant frequency ππΉ , as illustrated in Figure 8 top panel. The response curve of the center
neuron has a single peak at ππ ~ππΉ, as shown in Figure 8 bottom panel. But for a Rectangular
input function, the Fourier transform is a Sinc function with multiple peaks and troughs. Each
time a peak matches the critical frequency ππΉ in the Fourier space, the corresponding response
curve of the neuron at the center will have a local maximum at ππ ~ππΉ; i.e. the response curve
oscillates with stimulus size. Such oscillations were reported in previous experiments (Sengpiel
et al., 1997; Anderson et al., 2001). The population activity will also oscillate as a result of
resonance. When a peak of the input matches the critical frequency ππΉ in the Fourier space, the
population activity peaks at the center neuron. In contrast, when a trough of the input matches
the critical frequency ππΉ , the center neuron corresponds to a trough in population activity.
Therefore, as the input size increases, the network repeatedly gains and loses resonance; and the
population activity will alternate between oscillatory and non-oscillatory states.
Page 31
9. Summary
In this chapter, we studied a linear rate model for both excitatory and inhibitory neuronal
populations of a single preferred orientation, with a connectivity matrix dependent on the relative
position of the neurons. We searched for conditions when the connectivity functions produced
surround suppression in both excitatory and inhibitory neurons at the center of the stimulus. For
Gaussian input functions, we obtained analytic solutions with surround suppression boundary
conditions simplified from general surround suppression constraints. In addition, if the inhibitory
connections were short ranged, in order for the inhibitory response to decrease with increasing
stimulus size, two conditions must be met: (1) the network must be an ISN; and (2) the πΌ β πΈ
connections must be longer-ranged than the πΈ β πΈ connections.
We performed an exhaustive parameter space search to obtain general numerical solutions
without additional simplifications and approximations. The numerical solutions verified the
findings in the analytic results. We calculated the suppression index to characterize the strength
of the surround suppression for each numerical solution. Excluding solutions with insignificant
surround suppression (ππΌ < 5%), there were two key factors for a network to generate
significant surround suppression: (1) strong recurrence in the πΈ β πΈ sub-network (π ππ > 1, ISN
regime) and (2) wider πΌ β πΈ connection width than πΈ β πΈ connection width. In general, many
networks with π ππ > 1.5 and πππ > 2πππ showed surround suppression effects as long as the
network is overall stable.
Networks with strong surround suppression (ππΌ β₯ 50%) are generally less stable compared to
the ones with weaker ππΌ values. When the real parts of the leading eigenvalues of π (π)
approaches 1, the corresponding network behaves like a narrow band-pass filter around a critical
Page 32
filter frequency ππΉ . Approximation around this critical filter frequency provides a reasonable
estimate of the maximal responses for solutions with strong surround suppression. We also
predicted that the critical stimulus size ππΉ β‘ 1/ππΉ is generally smaller than the maximal
response stimulus size ππ .
Next, we introduced variability to the model, allowing sparse connections in the connectivity
matrix (200 to 800 connections per neuron in our simulations, instead of an all-to-all connection).
Depending on the population ππΌ of the dense matrix, the ππΌ values for individual neurons varied
widely in the sparse network. Neurons with very strong surround suppression were found in such
sparse networks, while such networks generally stayed away from instability. When the dense
network had relatively smaller population ππΌ, a unimodal distribution of ππΌ values was seen in
the sparse network. Furthermore, strong population ππΌ in the dense matrix led to a bimodal
distribution of the ππΌ values in the sparse network. Both distributions had been reported by
previous experiments. In contrast, if the corresponding dense matrix only produced insignificant
surround suppression (ππΌ < 5%), the ππΌ values of neurons in the sparse network remained small.
We also examined the effect of different input functions and different levels of input blurring. In
the results with Gaussian input functions and no blurring, there is a group of solutions with
strong ππΌ, whose maximal response stimulus sizes ππ could be predicted by the critical stimulus
size ππΉ . With input blurring, there were more surround suppression solutions in addition to the
ones obtained without input blurring. Among the solutions with strong surround suppression,
there was a new group of solutions corresponding to networks with globally inhibitory recurrent
inputs. In this group, the maximal response stimulus size ππ was small and was independent of
the critical stimulus size ππΉ . Meanwhile, Rectangular input function also increased the number of
Page 33
surround suppression solutions, and the solutions with strong ππΌ had similar distributions in the
parameter space compared to solutions with Gaussian input function.
In general, the response curve given by Equation 5 can also be interpreted from a signal
processing perspective where the input signal passes through the connectivity filter πΌ β
π (π) β1. For solutions with critical frequency ππΉ > 0, the population activity in the linear space
may oscillate due to resonance at ππΉ . This prediction should be tested by future experiments.
Page 34
Chapter III: Properties of spontaneous and sensory-driven
activities in the primary visual cortex of awake ferret
1. Experimental procedure and data acquisition
In this chapter we study extra-cellular recordings of awake ferrets (courtesy of Michael Weliky's
lab at University of Rochester). The recordings were gathered from linear arrays of 16 electrodes,
implanted in the primary visual cortex of ferrets. Each electrode array was placed at 300β500ΞΌm
depth in layer 2/3 of the primary visual cortex. The entire electrode array spanned 3mm with a
200ΞΌm distance between neighboring electrodes, covering a large portion of primary visual
cortex. In most cases, each electrode recorded multi-unit signals from neurons in the proximity.
A moving window average algorithm was used to remove background spiking activity and to
identify the spike trains (see Appendix B, Section 3 for detailed information about the
experimental procedure and data acquisition).
The experiment was performed in awake animals at different developmental stages. After birth,
eyes of the ferrets remained closed until about a month later. The first age group was at postnatal
age P29-P30 (n=3), around eye opening. The second group was at postnatal age P44-P45 (n=3),
during the critical period for ocular dominance plasticity, when orientation tuning and long-range
horizontal connections had developed. The third group of postnatal age P83-P86 (n=4) was the
typical matured age of the animal. A late group, 1 to 3 months after the third age group, was also
Page 35
included in the experiment at postnatal age P129-P168 (n=6). During the experiment, a given
animal only contributed to the recordings of a single age group.
The experiment also studied the effect of different viewing conditions. Spontaneous activities
were measured under complete darkness. Sensory-driven activities were measured with both a
natural scene movie and a white noise stimulus. The movie stimulus was taken from the movie
'Matrix', and so had spatial and temporal correlations typical of natural vision. Experiments on a
given day of each age group usually contained 15 sessions of 100 seconds recording for each
viewing condition, with the exception of P30a (26 sessions) and P168 (11 sessions). Recordings
of the same viewing condition were done within one recording block, with 5 seconds interval
between successive recording sessions.
2. Principal Component Analysis of the spike trains and the dominance of a
spatially long-ranged principal component
Recordings from the experiments in the previous section are 100 seconds long spike trains. We
received the data as spike counts in 2ms bins. The recording sessions were grouped by postnatal
ages and viewing conditions. Figure 14 shows examples of the spike trains in each age group
with all three viewing conditions. In the matured age groups P83-P86 and P129-P168,
microbursts across all electrodes can be seen in the spike train.
Fiser et al. (2004) previously examined this data, separately studying the spatial correlations and
temporal correlations. They found that: (1) the correspondence between sensory-driven activity
and the structure of the input signal was weak in young animals, but systematically improved
Page 36
with age; (2) correlations in spontaneous neural firing were only slightly modified by visual
stimulation; and (3) oscillations emerged in the spontaneous activities in the matured age group.
In this study, we examine the spatiotemporal structure of the recordings. We first re-sample the
spike trains into 10ms bins to reduce the number of empty bins. Then we normalize the re-
sampled data so that the recording on each electrode has zero mean and unit variance. To
identify spatially uncorrelated signals, we apply Principal Component Analysis (PCA) to the
normalized data for each recording session, analyzing the set of equal-time spatial vectors in
each 10ms bin, and obtained 16 uncorrelated spatial components. Figure 15a shows an example
of the PCA result from P129, dark viewing condition, session 1. The top panel illustrates the
shape of all 16 principal components, ordered by their contributions to the total variance (shown
in the bottom panel). The 1st principal component (1st PC) is a dominant mode, accounting for
more than 50% of the total variance. This mode is also long-ranged in space, covering the 3mm
span of the electrode array. Other PC modes contribute less to the total variance; and modes with
higher spatial frequencies tend to have less contribution to the variance.
In addition to the principal component analysis, we also examine the overall correlation structure
of the activity pattern by calculating the correlations between spike trains with time lag π and
spatial separation π₯, i.e. the spatiotemporal cross-covariance matrix:
πΆ π₯, π =1
ππ₯
1
2π ,π‘
πππ‘ππ+π₯
π‘+π + πππ‘ππβπ₯
π‘+π
(Eq. 10)
where π is the normalized spike train, π is the index of the electrode, and ππ₯ is a normalizing
factor (the number of electrode pairs separated by distance π₯). In Figure 15b, the top panel is the
Page 37
cross-covariance matrix calculated for P129, dark viewing condition, session 1. The
characteristic structure is approximately a spatially homogeneous mode spanning across all
electrodes, with a damped oscillation along the temporal axis. In the bottom panel, the
contribution of the1st PC mode is removed from the spike train, and the resulting plot only
contains the contributions from the other PC modes. In this plot, structure of the πΆ π₯, π is
characterized by a short-lived sharp peak around π = 0 and π₯ = 0, and all correlations quickly
decay to background level in about 50ms.
This spatially homogeneous 1st PC mode is a dominant component in all age groups, as shown
in Figure 15c, where recordings from the same age group are pooled together. The top panel
illustrates the averaged 1st PC mode in each age group, which resembles a spatial DC mode (the
dotted line: 0.25 on each of the 16 electrodes, i.e. a DC vector of unit length). The bottom panel
shows the percentage of the total variance of each corresponding PC mode, similar to the bottom
panel in Figure 15a. Table 1 below lists the absolute value of the difference between the 1st PC
mode and the DC mode on each electrode (normalized by the DC mode and then averaged over
all electrodes) and the percentage contribution of the 1st PC mode to the total variance:
Table 1
Age Group P29-P30 P44-P45 P83-P86 P129-P168
Difference (mean Β± sem) 12% Β± 2% 17% Β± 4% 9% Β± 2% 6% Β± 1%
% of Variance (mean Β± sem) 22% Β± 2% 23% Β± 1% 38% Β± 4% 42% Β± 2%
In all age groups, the 1st PC mode represents a spatially long-ranged component in the activity
pattern. Especially in the late age group (P129-P168), this mode is essentially a spatial DC mode.
The 1st PC also carries a bigger percentage of the total variance in matured age groups.
Page 38
We compare the auto-covariance curves of all PC modes and search for differences in their
temporal correlation structures. The 1st PC mode is a slow component for all age groups. As the
animal matures, the temporal correlations decay at a faster rate in all PC modes. However, the
reduction in temporal correlation is less in the 1st PC mode than in the other PC modes. Figure
15d shows the auto-covariance of the top four PC modes for age group P83-P86 (top panel) and
P129-P168 (bottom panel). The results are averaged across all recording days and all sessions for
a given age group. The 1st PC mode is a slow temporal mode, typically lasting for several
hundred milliseconds in the auto-covariance curve. Other PC modes decay into baseline level
much more quickly in matured age groups. We calculate the baseline correlation by averaging
the auto-covariance at large π (2s~5s) for each PC mode in each session. We define the
characteristic decay time of each PC mode to be the first time when the temporal correlation falls
within 2 standard error of the mean from the baseline level. The results of the spontaneous
activities are detailed in Table 2 (mean Β± sem, averaged within each age group, unit in ms). The
results under movie and noise viewing conditions can be found in Supplemental Table 1 and 2,
Appendix B Section 4a.
Table 2
Age Group P29-P30 P44-P45 P83-P86 P129-P168
PC 1 907 Β± 218 433 Β± 258 253 Β± 92 318 Β± 27
PC 2 743 Β± 203 373 Β± 303 50 Β± 7 58 Β± 14
PC 3 793 Β± 519 310 Β± 235 40 Β± 9 68 Β± 32
PC 4 453 Β± 179 237 Β± 79 45 Β± 12 55 Β± 14
PC 5 437 Β± 252 193 Β± 113 48 Β± 18 30 Β± 6
PC 6 280 Β± 135 153 Β± 84 48 Β± 9 37 Β± 3
PC 7 343 Β± 181 177 Β± 215 38 Β± 12 38 Β± 5
Page 39
PC 8 343 Β± 170 193 Β± 118 38 Β± 11 38 Β± 9
PC 9 290 Β± 142 190 Β± 125 45 Β± 12 38 Β± 5
PC 10 293 Β± 205 190 Β± 130 35 Β± 13 37 Β± 4
PC 11 333 Β± 189 167 Β± 122 33 Β± 13 38 Β± 7
PC 12 270 Β± 155 120 Β± 60 35 Β± 13 33 Β± 11
PC 13 210 Β± 162 67 Β± 3 33 Β± 13 28 Β± 5
PC 14 193 Β± 141 63 Β± 9 38 Β± 14 28 Β± 7
PC 15 147 Β± 99 63 Β± 3 35 Β± 13 25 Β± 3
PC 16 90 Β± 70 50 Β± 12 30 Β± 9 27 Β± 3
In general, PC modes with less contribution to the total variance tend to decay faster. In the
matured age groups P83-P86 and P129-P168, the 1st PC mode has significantly larger
characteristic decay time than the other PC modes. The characteristic decay time of the 1st PC
mode is about 5 to 10 times bigger than the largest characteristic decay time of the other PC
modes.
3. Development of spontaneous oscillation
In the previous section, we calculated the generic characteristic time for all PC modes,
independent of the shape of the auto-covariance curves. In this section, we focus on the PC
modes with distinct oscillation structures in the auto-covariance, and obtain the oscillation
frequencies and decay time constants by curve fitting.
The four age groups in the experiment cover important developmental stages of the primary
visual cortex. We examine the changes in the spontaneous activity by comparing the auto-
Page 40
covariance curve of the dominant 1st PC mode, as shown in Figure 16 top panel, averaged in
each group.
Upon eye opening (age group P29-P30), the auto-covariance shows long temporal correlations.
This long-ranged correlation becomes shorter in later age groups, in agreement with results of
previous studies (Chiu and Weliky, 2001; Fiser et al. 2004). In addition, the temporal structure of
the 1st PC has a strong tendency to oscillate in matured age groups. This spontaneous oscillation
usually disappears after several hundred milliseconds. To quantify both the fast and slow
components of the auto-covariance curve, we fit the experimental data with a model containing
three terms: a damped oscillation term, an exponential decay term and a constant term (baseline):
πΆ1 exp βπ‘
π1 cos 2πππ‘ + π0 + πΆ2 exp β
π‘
π2 + πΆ3
(Eq. 11)
To prevent over-fitting, we apply a nested model test in the curve fitting results for each day of
the experiment (Appendix B, Section 4b). The damped oscillation term is significant in P85,
P129, P134, P135 and P168 (Chi-square test, 5% significance level). Figure 16 bottom panel
shows an example of the curve fitting result in P129. The key parameters of the curve fitting
(data sets with significant oscillation only) are listed in Table 3. More detailed results can be
found in Supplemental Table 4, Appendix B Section 4b.
Table 3
Age P85 P129 P134 P135 P168
Page 41
π (Hz) 12.2 13.6 13.8 8.57 10.7
ππ(ms) 26.2 77.2 96.0 94.1 72.1
ππ(ms) 131 326 1015 221 347
The damped oscillation is a fast component (π1<100ms), while the exponential decay term is a
slow component (π2>100ms). The oscillation frequency is in the range of 8~14Hz, comparable
to the πΌ brain wave (8~14Hz).
4. Spontaneous oscillation in networks with surround suppression
In this section, we examine network solutions obtained in Chapter II that showed surround
suppression effects, and search for the ones that would generate spontaneous oscillation. We
study the network as a Hebbian assembly, and estimate the decay time and oscillation frequency
in a Hebbian amplification scenario. The predictions are compared with the damped oscillation
observed in the matured animals from the previous section.
As mentioned in Section 1, experiments in this chapter were done on a linear electrode array.
Thus we study a 1-dimensional rate model with both excitatory and inhibitory populations, given
by Equation 2, Chapter II. Under complete darkness, the primary visual cortex receives
spontaneous inputs from LGN. We assume that the mean of the LGN spontaneous activity is
stationary in the time scale of 1s. For simplicity, we first assume the fluctuation around this mean
activity is white noise. Effects of the temporal correlations in the LGN inputs are discussed at the
end of this section. The damped oscillation (π1<100ms, Table 3) discussed in Section 3 is fast
Page 42
compared to the 1s time scale, and its dynamics can be approximated as perturbations around the
fixed point given by the LGN mean activity. The slow component in Section 3 is comparable or
even larger than this time scale, and is not modeled here.
In cortex, each neuron only projects one type of synapse (either excitatory or inhibitory) to other
neurons depending on its cell type. In particular, the sign of the inhibitory-to-excitatory
connection is opposite to that of the excitatory-to-inhibitory connection. So the connectivity
matrix of the recurrent network is not symmetric and is non-normal, meaning its eigenvectors are
not mutually orthogonal (properties of non-normal matrices are detailed Appendix B, Section 5b).
For simplicity, we study the dominant 1st PC as an effect of Hebbian amplification (Chapter I,
Section 2), where the pattern with the slowest decay rate (determined by the largest real part of
the eigenvalue) will emerge from the spontaneous activity by accumulating to the highest
amplitude. Predictions of the Hebbian amplification are compared with the experimental results
at the end of this section.
The eigenvalue of the weight matrix at spatial frequency π is given by Equation 12 (See
Appendix B, Section 2a for the mathematical details):
π(π) =Tr(π (π))
2 Β±
Tr(π (π))
2
2
β det π (π)
(Eq. 12)
where π (π) is the weight matrix in the Fourier space given by Equation 6 in Chapter II. The
network activity oscillates when the eigenvalue has an imaginary part, and the oscillation
frequencies is given by π =ππππ π π
2πππ, where ππ is the membrane time constant.
Page 43
The 1st principal component corresponds to a spatially homogeneous mode and can be
approximated by a spatial DC mode (π = 0). Figure 17 top panel is the histogram of the
predicted oscillation frequency of the spatial DC mode (π = 0), for all networks showing
surround suppression effect with Gaussian input and without blurring (Chapter II, Section 4),
assuming reasonable membrane time constants ππ = ππ = 10ππ . About 85% (6121 out of 7199)
of such networks show oscillations in the spatial DC mode, with a peak around 12Hz (mean Β± std:
12.1 Β± 5.3 Hz). About 37% (2646 out of 7199) of all solutions show oscillations in the 8~14Hz
range, i.e. the range of spontaneous oscillation observed in the matured age groups (Table 3).
In Section 2, we compared the auto-covariance curves of different PC modes, and computed the
characteristic decay time for each PC mode, regardless of the shape of its auto-covariance curve.
In matured age groups, the 1st PC mode is much slower compared to the other PC modes. The
characteristic decay time of the 1st PC mode is about 5 to 10 times bigger than the largest of the
other PC modes. We use the spatial DC mode (π = 0) to approximate the 1st PC mode; and for
the other PC modes, we use a Fourier mode at frequency π = 2π/3 (mmβ1) as an estimate,
which is the lowest non-zero spatial frequency measured from the 3mm span electrode array.
The time constant of Hebbian amplification is given by ππ» =πm
1βmax (ππππ (ππ )) , where ππ is the
membrane time constant. We search the surround suppression results (with Gaussian input and
no blurring) for networks whose ratio of ππ»π·πΆ/ππ»3ππ
is within 5 to 10, as shown in Figure 17
bottom panel. Within this range (the highlighted area), we have 0.825 < πππ₯ ππππ ππ·πΆ <
0.925, and ππ»π·πΆ is about 57~133ms. This prediction is in agreement with the decay time
constant of the damped oscillation in age group P129-P168 discussed in Section 3. Further
Page 44
improvements of this estimate would probably require the non-normal effects to be incorporated
into the model.
The connectivity matrix at a given spatial frequency can be reduced to a 2-by-2 matrix in the
Fourier space (Equation 6 in Chapter II). Damped oscillation in the auto-covariance of the DC
mode indicates that the eigenvalues of the 2-by-2 matrix are complex conjugates. For the spatial
DC mode, the trace of the weight matrix is: Tr π π = 0 = π ππ β π ππ = 2real ππ·πΆ , so the
network must function in the ISN regime (π ππ > 1) when real ππ·πΆ > 0.5. The damped
oscillation results in age group P129-P168 have π1 > 72ππ . LGN inputs typically have
correlation time around 50ππ (Wolfe and Palmer, 1998). Therefore, the Hebbian amplification
contributed by the cortical network gives ππ» > 22ππ , i.e. real ππ·πΆ > 0.545 if we assume
reasonable membrane time constants of 10ms. Thus, the networks measured in the experiments
correspond to inhibition stabilized networks in our model.
5. Modulations of the auto-covariance by sensory stimuli
In Section 3 and Section 4, we studied the auto-covariance results of the spontaneous activities.
In this section, we focus on the effects of the sensory stimuli. As mentioned in Section 2, the 1st
PC emerged as a spatially homogeneous and temporally slow mode in sensory-driven activities
(more details in Appendix B, Section 4a). Sensory stimuli significantly increase the firing rate in
early age groups P29-P30 and P44-P45 (increase in firing rate: 56% for P29-P30 Movie, 43% for
P44-P45 Movie, 48% for P29-P30 Noise, and 52% for P44-P45 Noise). In matured age groups
P83-P86 and P129-P168, the spontaneous activity becomes stronger and the sensory stimuli only
Page 45
increase the firing rate by a smaller amount (increase in firing rate: 6% for P83-P86 Movie, 2%
for P129-P168 Movie, 25% for P83-P86 Noise, 13% for P129-P168 Noise).
Figure 18(a-d) contains the auto-covariance of the 1st PC mode in all age groups under Dark,
Movie and Noise viewing conditions. In age group P29-P30 (Figure 18a), around eye opening,
the cortex just starts to receive sensory stimuli. The temporal structures of the auto-covariance
curves are similar in all 3 viewing conditions, with long temporal correlations, suggesting the
immature animals lack the ability to rapidly respond to visual stimuli. In age group P44-P45
(Figure 18b) and P83-P86 (Figure 18c), natural scene Movie stimuli have a tendency to increase
the decay time of the 1st PC mode. In age group P129-P168 (Figure 18d), the modulation effect
of Movie stimuli is very small, giving almost identical auto-covariance structure under Movie
and Dark viewing conditions. Significant oscillations are observed in the Movie response as well
as in the spontaneous activity. In general, Noise stimuli did not have a strong modulation effect
as seen with Movie stimuli, except in the late age group of P129-P168, where a strong and
sustained oscillation (~10Hz) can be induced by the Noise stimuli. In P129, P134 and P142, the
oscillation lasts for over 2 seconds in the auto-covariance curve. Oscillations in P135, P151 and
P168 are weaker in amplitude and shorter in time span. In the case of strong oscillation,
synchronized bursts dominate the spike train, as shown in the raster plot of Figure 16d (bottom
graph, P129 noise, Session 1). To compare the impact of various stimuli in each age group, we
calculate the difference in the auto-covariance curves (as a vector norm) divided by half of the
vector norm of the sum. The results are shown in Figure 18e, averaged within each age group.
The differences between spontaneous and sensory-driven activities are small in the eye-opening
group P29-P30, and become larger as the animal matures. In late matured group P129-P168, the
difference between Dark and Movie viewing conditions becomes smaller, but is still larger than
Page 46
that of group P29-P30. The difference between Dark and Noise viewing conditions becomes
larger due to the sustained oscillation induced by Noise stimuli.
We perform a least squares fit on the sensory-driven activities of the age group P129-P168, using
the same model given by Equation 11, Section 3. An example of P129 Movie is shown in Figure
19 top panel. The oscillation frequencies on different recording days with different view
conditions are listed in Table 4 (all results show significant oscillations, chi-square test, 5%
significance level).
Table 4
Age P129 P134 P135 P142 P151 P168
Movie (Hz) 12.6 14.0 10.2 11.5 12.8 14.6
Noise (Hz) 9.9 10.0 11.0 10.0 10.2 15.0
Figure 19 bottom panel shows the power spectra under all 3 viewing conditions for P129. The
bump is seen under all conditions, and corresponds to oscillations around 8~14 Hz. The Noise
stimuli generate strong and long-lasting oscillations (the strong and sharp peak around 10Hz in
the figure). The second harmonic of this oscillation can be seen around 20Hz.
6. Absence of orientation map structure in both spontaneous and sensory-
driven activities
In Figure 15a, all PC modes are ordered by their contributions to the total variance. These PC
modes resemble Fourier modes with increasing spatial frequency. Since the Fourier modes also
Page 47
form an orthogonal set of bases, we project the recording on a set of 16 Fourier bases and
calculate the variance in each Fourier mode. In Figure 20 (a-d), we plot the variance vs. the
spatial frequency of each Fourier mode in log-log scale (excluding the π = 0 mode), for all
viewing conditions and all age groups. In each panel the dotted lines are given by the linear least
squares fit, color coded according to the viewing conditions. The curve fit parameters are given
in Table 5:
Table 5
Slope Intercept ππ
P29-P30 Dark -0.482 9.07 0.977
P29-P30 Movie -0.460 9.04 0.961
P29-P30 Noise -0.446 9.07 0.967
P44-P45 Dark -0.430 9.07 0.983
P44-P45 Movie -0.455 9.02 0.987
P44-P45 Noise -0.477 9.01 0.979
P83-P86 Dark -0.427 8.78 0.996
P83-P86 Movie -0.418 8.70 0.996
P83-P86 Noise -0.492 8.66 0.990
P129-P168 Dark -0.586 8.75 0.986
P129-P168 Movie -0.623 8.64 0.986
Page 48
P129-P168 Noise -0.695 8.60 0.979
The variances of the Fourier modes in the log-log plot are linearly dependent on the spatial
frequencies, i.e. the power of the network activity decays as a negative power of the frequency.
This roughly power law dependency suggests that the network activity in the awake state has no
characteristic length and is roughly scale-invariant. In the primary visual cortex, neurons are
organized in orientation columns, where neurons with similar preferred orientations are clustered
in small patches (the columnar organization of the visual cortex is further explained in Appendix
A, Section 3). The orientation map is usually measured in anaesthetized animals with a bar-
shaped stimulus. The preferred orientations repeat periodically across cortex, with a typical
period in the range of about 0.5~1mm (Chapman et al., 1996; Rao et al., 1997).
This characteristic length in the orientation map is absent in the roughly power law dependency
in Figure 20. In the experiment, the 16-electrode array was randomly placed over the primary
visual cortex. To test if this characteristic length is obscured by the coarse sampling of the
electrode array across irregularly shaped orientation columns, we simulate the same recording
process as in the experiment, over an orientation maps measured from P42 ferret (Chapman et al.,
1996). In the simulation, the measured orientation map is discretized into a pixel map with the
resolution of 10Β΅m. 6161 excitatory neurons and 6161 inhibitory neurons are placed on a 3mm
by 5mm grid over the orientation map, with a 50Β΅m spacing between neighboring neurons. The
preferred orientation of each neuron is given by the underlying pixilated orientation map. The
weight matrix depends on both the spatial location and the preferred orientation of the neurons.
We assume the spatial dependency of the weight matrix to be a Gaussian function as in Chapter
Page 49
II, and the contribution of the orientation tuning is another Gaussian function of the difference in
the preferred orientation, with the tuning width ππ . Then, the weight matrix is given by:
π π₯ 1, π₯ 2, π1, π2 ~exp β π₯ 1 β π₯ 2 2
2ππ₯2
exp β π1 β π2 2
2ππ2
(Eq. 13)
We simulate the spontaneous activity by using an input sequence with a slow varying mean plus
white noise fluctuations, similar to the LGN inputs discussed in Section 4. The mean inputs are
drawn from random spatial patterns over the 16 electrode array. We assume the LGN mean input
is stationary in the time scale of 1s, and these spatial patterns are refreshed at 1s interval. The
sequence of these patterns is convolved with a Gaussian temporal kernel (of 1s width) to
generate a continuously varying mean input.
We choose 16 recording sites with the same configuration as in the 16-electrode array, and
record the activities of the simulated network. The recording starts 10 seconds after the onset of
the input. In the experiment, the 16-electrode array did not distinguish excitatory and inhibitory
neurons. Thus, in our simulation, the recording on each sampling location is a weighted average
of the firing rates of the surrounding neurons. The weight function for averaging is a Gaussian
function of the distance between the sampling location and the surrounding neurons. The width
of this Gaussian function is 25Β΅m.
We first simulate a reduced version of the model with excitatory neurons only. Recordings from
the simulations are normalized and processed in the same way as the experimental data. We plot
the variance vs. the spatial frequency of the corresponding Fourier mode for the simulated data in
a log-log scale as shown in Figure 21 top panel, color coded by different spatial connection
Page 50
widths, covering a wide range: ππ₯ = 0.1, 0.2, 0.5, 1.0, 2.0 mm. The roughly power law
dependency is absent in this simulation. Conversely, for ππ₯ > 0.2mm (very small ππ₯ are
biologically unlikely), the simulation results show a peak around π = 1/750mβ1,
corresponding to the typical length scale for orientation columns.
The simulation results of the full model with both excitatory and inhibitory neurons are shown in
Figure 21 bottom panel: an example of a deterministic network with connectivity parameters
same as Population 1 in Chapter II Section 6 (πππ = 1.0, πππ = 0.5, πππ = 1.9, πππ = 0.3,
πππ = 1.5, πππ = 1.0, πππ = 1.25, πππ = 1.0 and ππ = π/8). The dashed line is the linear least
squares fit of the data. The fitting results are: slope = β0.200, intercept = β2.767 and
π2 = 0.968. The simulated data also roughly have a power law dependency over the spatial
frequency, though with considerably smaller negative exponent than in the experimental data.
The characteristic length corresponding to orientation columns is absent, similar to the
experimental data.
The differences between these simulations suggest that the inhibitory population is necessary for
the model to reproduce the experimental data.
7. Summary
We analyzed the recordings of both spontaneous and sensory-driven activities in the primary
visual cortex of awake ferrets. Using Principal Component Analysis, we first analyzed the
characteristics of the principal components in the spontaneous activity. In all ages, the first
principal component was responsible for a large part of the total variance compared to the other
PC modes. This 1st PC mode was spatially homogeneous, and in matured age groups, it was
Page 51
effectively a long-ranged spatial DC mode. The 1st PC mode also showed a tendency to oscillate
with increasing animal age. In the power spectrum of matured age groups, there was a peak in
the 8-14Hz range, corresponding to Ξ± band brain waves. Similar oscillations were also seen in
the activities with movie and noise stimuli in matured age groups. With noise stimuli,
particularly in the late age group of P129-P168, some oscillations became very strong and lasted
for over 2 seconds in the auto-covariance curves, and a synchronized burst of about 10Hz could
be seen across all electrodes in the raster plot of the spike train.
We also examined modulations of the activity patterns by sensory stimuli. The natural movie
stimuli contained spatially and temporally correlated information, and increased the temporal
correlation of the 1st PC mode in most age groups. In the late age group P129-P168, the
difference in auto-covariance curves between the dark and the movie viewing conditions became
much smaller. This suggested that the spontaneous activity has become biased toward processing
retinotopically correlated inputs (Berkes et al., 2011). The shapes of auto-covariance curves were
similar between dark and noise viewing conditions in early age groups, and became different in
late age groups due to the strong and sustained oscillation.
We used a model with both excitatory and inhibitory neurons (same as in Chapter II) to study the
spontaneous activities. In matured age groups, the spontaneous oscillation of the 1st PC mode
can be fitted by a damped oscillation. We applied network parameters from the surround
suppression results, and assumed a Hebbian amplification scenario, where the real part of the
leading eigenvalue gave an estimate of the speed of the decay, and the oscillation of the 1st PC
mode arose from the imaginary part of the corresponding eigenvalue. The predictions of the
model were within the range of the experimental observations. Furthermore, the model predicted
Page 52
that when the auto-covariance showed a significant damped oscillation in age group P129-P168,
the network functioned in the ISN regime.
In the primary visual cortex, neurons are organized in orientation columns. To examine the
spatial structure of both spontaneous and sensory-driven activities, we projected the activities on
a set of spatial Fourier modes and calculated the corresponding variance at different spatial
frequencies. The result showed a roughly power law dependency without any characteristic
length. This roughly power law dependency was reproduced by simulations in networks with
both excitatory and inhibitory neurons on a measured orientation map. In contrast, the
characteristic length for orientation columns appeared in simulations of networks on the same
orientation map but with excitatory neurons only.
Page 53
Chapter IV: Conclusions and Discussions
In the previous chapters, we studied a linear rate model for networks with both excitatory and
inhibitory neurons. We applied this model to two separate studies in the primary visual cortex: (1)
the conditions to achieve surround suppression, and (2) the properties of the spontaneous and
sensory-driven activities. In both studies we constructed the connectivity weight matrix with
biologically reasonable parameters. In the surround suppression study, we modeled a network of
neurons sharing the same preferred orientation with a weight matrix where the connectivity
weights between neurons depend only on their relative locations. The same model was extended
in the study of spontaneous and sensory-driven activities to include an orientation-dependent
component to the weight matrix. In both studies, we obtained parameters of the weight matrix in
which the excitatory sub-network was unstable by itself, and inhibitory neurons were needed to
stabilize the entire network, i.e. the network was inhibition stabilized.
In the study on surround suppression effects, we used two sets of input functions: the Gaussian
function and the Rectangular function. These functions, with different levels of input blurring
along the visual pathway, simulated stimuli with different shapes on the edge. The surround
suppression effects were modeled by applying different constraints to the responses of the
neurons at the center of the stimuli. We showed both analytically and numerically that the
network must be an ISN if inhibition was short in range. More generally, we searched for
numerical solutions given by these constraints in a wide range of connectivity parameters. We
Page 54
identified several groups of solutions from results with different input functions and different
blurring widths:
The first group represents solutions in the ISN regime. Solutions in this group can lead to strong
surround suppression (ππΌ β₯ 50%), and such solutions can be seen in all combinations of input
conditions (both types of input functions, with or without input blurring). In this group,
excitatory to inhibitory connections are longer in range than excitatory to excitatory connections.
Such solutions are characterized by a critical filter frequency ππΉ; and surround suppression
effects can be attributed to the resonance at this critical frequency. In this scenario, when the size
of the stimulus matches the optimum size determined by the critical frequency, the response of
the center neuron reaches the resonance maximum. Further increase in stimulus size leads to loss
in resonance, i.e. the neuronal response is effectively suppressed by larger stimuli. When
variability is introduced to the model (in terms of sparse connections in the weight matrix), we
are able to generate 2 types of distributions of the population ππΌ, similar to those reported by
previous experiments (Walker et al., 2000; Jones et al., 2000; Akasaki et al., 2002). Furthermore,
surround stimuli induce a transient increase in the inhibitory conductance followed by a steady-
state decrease in inhibitory conductance in the surround-suppressed steady state (Ozeki et al.,
2009). The transient increase and the steady-state decrease are typical behaviors of these ISN
solutions and are not expected for non-ISN solutions, which should simply show a monotonic
increase in inhibition received.
The second group of solutions represents networks where the recurrent inputs to the central
neuron are globally inhibitory. These solutions are generally non-ISN solutions, and many of
such solutions also show strong surround suppression (ππΌ β₯ 50%). Without input blurring, these
solutions generate monotonically decreasing response curves, and thus do not qualify as
Page 55
surround suppression solutions. With input blurring, the feed-forward input received by the
center neuron is the stimulus convolved with the blurring function. When the stimulus size is
much smaller than the width of the input blurring, the convolution is roughly proportional to the
sum of the stimulus over positions. Thus, the response curve of the center neuron always shows
an initial summation. As the stimulus size becomes larger, the feed-forward input saturates and
the recurrent inhibitory effect leads to decreasing response with increasing stimulus size, i.e.
surround suppression.
The summation sizes of such solutions are comparable to the width of the blurring, which is
typically smaller than the width of lateral πΈ β πΈ connections. In macaque monkeys, typical size
of the summation surround is about 1o within 2 β 8o from the fovea (Angelucci et al., 2002).
The magnification factor (the change in cortical position corresponding to a given change in
retinotopic position) is about 2.3 mm/deg at 5o from the fovea (Van Essen et al., 1984). Thus, to
match the experimental measurements, the width of the blurring should cover ~2mm of cortex.
Along the visual pathway, neighboring LGN cells have spatially overlapping receptive fields
(Hubel and Wiesel, 1962; Hammond 1972); LGN projections spreads over about 1mm of cortex
and cortical dendrites extend several hundred microns (Salin et al., 1989; Yamamoto et al., 1989).
Putting all these effects together, large input blurring is possible but might be difficult to achieve.
The third group of solutions appears only in results with Rectangular input functions. These
solutions generally have very short ranged πΈ β πΌ connections and show insignificant surround
suppression (ππΌ < 5%). With Gaussian input functions, such solutions generate monotonically
increasing response curves. In contrast, the Fourier transform of a Rectangular input function is a
Sinc function with a large central peak, and the area under the central peak is larger than the
integral of the entire function in the Fourier space (corresponding to the response at input size
Page 56
π β β). Thus, the global peak in the response curve may appear when the central peak of the
Sinc function is optimal for the connectivity filter πΌ β π (π) β1
. Such solutions become
surround suppression solutions as defined in Constraint A, Chapter II Section 2.
Both the first and the second group of solutions can lead to strong spatial oscillations in
population activity patterns. The ISN solutions with strong surround suppression (ππΌ β₯ 50%) are
characterized by a critical filter frequency ππΉ . The population activity of the network is given by
the inverse Fourier transform of the product of the inputs and the connectivity filter πΌ β
π (π) β1 around ππΉ . Therefore, when the size of the input is comparable to the critical stimulus
size ππΉ (given by 1/ππΉ), spatial population oscillations may emerge due to this resonance.
For Gaussian input functions, there is one global critical stimulus size ππΉ , corresponding to the
resonant frequency. As the size of the input increases, more neurons in the network are activated.
The period of the population oscillation increases, while the amplitude of the oscillation
decreases. At very large input sizes, the population activity becomes translation-invariant and the
oscillation disappears. In contrast, for a Rectangular input function, the Fourier transform is a
Sinc function with multiple peaks. The response curve of the neuron at the center can have
multiple local maxima due to resonance at these peaks in the Fourier space. Such oscillations in
the response curves were reported in previous experiments (Sengpiel et al., 1997; Anderson et al.,
2001). In addition, the population activity will also oscillate as a result of resonance. As the input
size increases, the resonance peak of the connectivity matrix scans across the peaks and troughs
of the input in the Fourier space. The network gains resonance at each local maxima and minima,
and loses resonance in between. Thus, the population activity will alternate between oscillatory
and non-oscillatory states.
Page 57
Solutions in the second group represent networks with globally inhibitory recurrent connections.
When the stimulus has sharp edges (e.g. Rectangular input functions), such solutions can
generate population oscillations by a chained lateral inhibition mechanism (Adini et al., 1997):
neurons near the edges of the stimulus receive partial surround inputs and are less surround-
suppressed, creating a peak in the population activity. This peak suppresses nearby neurons,
while neurons further away are facilitated due to a decrease in suppression from their immediate
neighbors. Such chained lateral inhibition generates a standing wave in the population activity
that peaks near the edges of the stimulus. Depending on the specific stimulus size, the neuron at
the center can be at different phases in the standing wave. As a result, the response curve of the
center neuron will also oscillate with stimulus size.
Both mechanisms mentioned above lead to oscillatory response curves. However, the population
activity patterns are different between these two mechanisms: resonant oscillation predicts that
the population activity will alternate between oscillatory and non-oscillatory states, while the
population activity shows constant oscillation in the chained lateral inhibition scenario. When the
connectivity matrix has a sharp πΏ-function like peak, the resonant oscillation dominates the
population activity. Otherwise, the connectivity matrix is 'aware' of the spatial structure of the
input, and for Rectangular input functions, population oscillations can be a mixture of these two
mechanisms.
In the study of spontaneous and sensory-driven activities, we analyzed the multi-unit recordings
from the primary visual cortex of awake ferrets. Principal Component Analysis of the recorded
Page 58
data showed that the principal component with the largest contribution to the total variance (the
1st PC mode) is a spatially homogeneous mode (the DC mode). The temporal correlation of the
1st PC mode became dominant and showed a tendency to oscillate (8-14Hz) with increasing
animal age. In the matured age groups, the auto-covariance curve of the 1st PC mode was
approximately a damped oscillation.
We studied spontaneous activities in a network of both excitatory and inhibitory neurons, with
parameters obtained from the surround suppression study. We studied the network as a Hebbian
assembly, where the real part of the leading eigenvalue gave an estimate of the speed of the
decay and the oscillation of the 1st PC mode arose from the imaginary part. The predictions of
the model were within the range of the experimental observations. The connectivity matrix at a
given spatial frequency can be reduced to a 2-by-2 matrix in the Fourier space. For a spatial DC
mode with damped oscillation, the strength of the πΈ β πΈ connection is given by: π ππ =
Tr π π = 0 + π ππ = 2real ππ·πΆ + π ππ , where Tr π is the trace of the matrix. Thus the
network functions in the ISN regime if real ππ·πΆ > 0.5. Curve fitting results of the damped
oscillation in age group P129-P168 indicated that real ππ·πΆ > 0.545 in the Hebbian
amplification scenario, given typical correlation time of 50ms in the LGN inputs and the
assumption of a membrane time constant of 10ms. Therefore, such networks measured in the
experiment correspond to inhibition stabilized networks in our model.
Both correlated (natural scene movie) and white noise stimuli were used in the experiment. At
eye opening, the cortex is new to external stimuli; and both types of stimuli induce trivial
changes in the correlation structure of the activity pattern. During the critical period, movie
inputs significantly increase the correlation time. This may be important for the development of
Page 59
the long-range horizontal connections and the orientation tuning in the primary visual cortex. In
matured age groups, correlation time becomes smaller for both spontaneous and movie-driven
activities, favorable for fast signal processing. In addition, modulations by movie stimuli become
small in age group P129-P168. This indicates a strong correspondence between the spontaneous
activity and cortical representation of natural sensory signals. Such correspondence suggests that
the spontaneous activities reflect the intrinsic dynamics of neural networks (Kenet et al., 2003;
Fiser et al. 2004; Berkes et al., 2011).
In the late age group P129-P168, noise stimuli can induce a strong and sustained oscillation that
lasts for more than 2 seconds in the auto-covariance curve. A synchronized burst (~10Hz) across
all electrodes emerges from the recordings with noise stimuli, and can be seen directly in the
raster plot of the spike trains. Oscillations in 8-14Hz correspond to Ξ± band brain waves. -band
activities usually appear when the visual cortex is in an idle state. The strong 10Hz oscillation
under noise stimuli also falls in the -band. Previous study (Kelly et al., 2006) suggested that
such oscillation may be attributed to suppression of competing distractions. The level of the
attention of the animal was not monitored in this experiment. Future experiments registering the
level of the attention would shed more light on these phenomena.
Fourier analysis of the activities of awake animals showed a roughly power law dependency
between the variance and the spatial frequency. This roughly power law dependency was
reproduced, on a measured orientation map, by simulations of networks with both excitatory and
inhibitory neurons. In the primary visual cortex, neurons are organized in orientation columns.
The characteristic length of this columnar structure is absent in the data; and the discrepancy
may be attributed to different states of arousal. Orientation-map-like structures have been
reported by previous study on spontaneous activities in anaesthetized animals (Kenet et al.,
Page 60
2003), however, the experiment examined in our study was done on awake animals. In the
anaesthetized state, the inhibitory connections are much stronger than that in the awake state.
The arousal states were not accounted for in the current model; and modifications in the
connectivity matrix may be necessary in future studies to model both anaesthetized and awake
states. One possible modification is to include a constant term to the orientation tuning
component of the connectivity matrix, representing an un-tuned component for all orientations.
In the awake state, the orientation map may be obscured by this un-tuned component. When
anaesthetized, inhibition reduces baseline activity and may suppress the contribution from this
constant term. Thus, the remaining orientation-tuned component could emerge more strongly in
the anesthetized state. This alternative hypothesis would require modifications in the network
circuitry of the current model.
In conclusion, we studied two seemingly unrelated phenomena: the surround suppression effect
and the spontaneous and sensory-driven activities in the primary visual cortex. The inhibition-
stabilized network model seems necessary to explain surround suppression effects, and is
consistent with the effects seen in spontaneous and sensory-driven activities. This suggests that
the ISN mechanism might play an important role in the neural circuitry in the primary visual
cortex.
Page 62
Figure 1. Typical stimulus configurations for surround suppression experiments
Typical stimulus configurations for surround suppression experiments. The center stimulus is a
drifting grating with optimal parameters (orientation, spatial/temporal frequency, etc) in the
classical receptive field. Surround suppression effect is the strongest when the surround stimuli
have parameters similar to that of the optimal stimulus in the classical receptive field.
Page 63
Figure 2. Mechanisms of the Difference of Gaussian (DoG) model
Mechanisms of the Difference of Gaussian (DoG) model. Top panel: the recurrent connections
from the center and the summation surround of the receptive field are modeled as an excitatory
Gaussian function (blue curve), while connections from the suppressive surround are modeled as
a broader but weaker inhibitory Gaussian function (red curve). The overall connectivity is given
by the difference of the two Gaussian functions, shaped like a 'Mexican hat' (black curve).
Page 64
Bottom panel: the inputs to the center neuron are calculated by integrating the effective
connectivity with the stimulus. Both the excitatory and the inhibitory inputs increase with the
stimulus size. The wide inhibitory input is responsible the surround suppression effect at large
stimulus sizes.
Page 65
Figure 3. The Inhibition Stabilized Network (ISN) model
Panel a: both excitatory and inhibitory neuronal response functions are assumed to be
generalized logistic function, as mentioned in Chapter I Section 3, where πΎ = 1.0, π = 0.5,
π΅ = 1.5, π = 7.0, π = 3.0 and π£ = 0.5. Panel b: stability of non-ISN. When the excitatory to
excitatory connection is weak (network parameters: πππ = 0.15, πππ = 0.7, πππ = 2.0, πππ =
Page 66
1.0, ππ = 0.7, ππ = 0.0), the excitatory nullcline has negative slope and the inhibitory nullcline
has positive slope at the fixed point. Panel d: increasing inputs (ππ = 0.15) to the inhibitory
population will shift the fixed point, creating a decrease in the firing rate of the excitatory
population and an increase in the firing rate of the inhibitory population. Panel c: stability of ISN.
When the excitatory to excitatory connection is strong enough (network parameters: πππ = 0.75,
πππ = 0.4, πππ = 1.7, πππ = 0.75, ππ = 0.3, ππ = 0.0), the excitatory nullcline has a segment of
positive slope (shown in dashed line). When the fixed point occurs along this segment, the
excitatory sub-network is unstable by itself. However, with recurrent inhibition, the fixed point
of the network can remain stable, i.e. the network is stabilized by inhibition. Panel e: In the ISN
scenario, increased inputs to the inhibitory population will lead to a new fixed point with lower
firing rate for both excitatory and inhibitory populations. In general, this additional input inhibits
the excitatory activity, and the network shifts to a less active state due to withdrawal of
excitation.
Page 67
Figure 4. π subspace of numerical solutions with Gaussian input function
Figure 4a. Numerical solutions for local inhibition with Gaussian input function (745
solutions).Search parameters: πππ = 1, πππ = (0.7, 0.9, 1.1, 1.3, 1.5, 1.7, 1.9, 2.1, 2.3, 2.5, 2.7,
2.9, 3.1, 3.3, 3.5, 3.7, 3.9), πππ , πππ = 0.05, πππ , πππ = (0.20, 0.35, 0.50, 0.65, 0.80), πππ , πππ =
(0.1, 0.2, 0.4, 0.8, 1.6, 3.2, 6.4), ππ/ππ = 1.0 and π0 = 0. Each disk represents surround suppression
solutions meeting the Constraints A, B and C in Chapter II Section 2, color coded by the number
of combinations of other parameters searched: πππ , πππ and πππ . All solutions have π ππ =
2πππππππ > 1 and therefore the corresponding networks function in the ISN regime. All
solutions are above the dashed curve given by πππ = π ππ
π ππ β1
1/2
, as predicted by the analytic
results in Chapter II Section 3.
Page 68
Figure 4b. 3-dimensional π subspace of numerical solutions. Parameters are specified in Chapter
II, Section 4 (Gaussian input function, ππ/ππ = 1.0 and π0 = 0, 7199 solutions total), color coded
by solution density, i.e. the number of combinations of other parameters searched: πππ , πππ , πππ
and πππ . The solution density is higher when πππ and πππ become large.
Page 69
Figure 4c. Same as Figure 4b, color coded by the averaged suppression index of the excitatory
response for all surround suppression solutions (averaged over the combinations of the other
parameters: πππ , πππ , πππ and πππ ).
Page 70
Figure 4d. Same as Figure 4b, color coded by the averaged suppression index of the inhibitory
response.
Page 71
Figure 4e. Projection of Figure 4b on the πππ vs. πππ plane. The solid curve is given by πππ2 +
πππ2 = πππ
2 = 1, where πππ2 + πππ
2 represents the effective width of the lateral inhibition by an
πΈ β πΌ β πΈ connection chain. All solutions satisfy πππ2 + πππ
2 > πππ2 . The region above the broken
line is given by the biological constraint πππ > πππ . In this region, the πΌ β πΈ projection is wider
than the πΈ β πΈ projection, i.e. πππ > πππ .
Page 72
Figure 5. π subspace of numerical solutions with Gaussian input function
Figure 5a. 3-dimensional π subspace of numerical solutions with parameters specified in
Chapter II, Section 4 (Gaussian input function, ππ/ππ = 1.0 and π0 = 0). All amplitudes are
normalized by the amplitude of the πΈ β πΈ connection in the Fourier space, color coded by
solution density. The plane is the least squares fit given by π ππ = 2.10π ππ + 0.35π ππ β
1.21π ππ .
Page 73
Figure 5b. Same as Figure 5a, color coded by the averaged suppression index of the excitatory
response.
Page 74
Figure 5c. Same as Figure 5a, color coded by the averaged suppression index of the inhibitory
response.
Page 75
Figure 6. Histogram of the amplitude of the π¬ β π¬ connection
Histogram of the amplitude of the πΈ β πΈ connection π ππ of solutions filtered by additional
constraint on the minimal ππΌ value (7199 solutions total, 3351 solutions for ππΌ β₯ 5%, 2237
solutions for ππΌ β₯ 10%, 1249 solutions for ππΌ β₯ 20% and 286 solutions ππΌ β₯ 50%). π ππ
determines whether the network is an ISN or not. Groups with stronger minimal ππΌ (ππΌ β₯ 10%,
ππΌ β₯ 20% and ππΌ β₯ 50%) are ISN solutions with very strong recurrent connections in the
excitatory sub-network, i.e. π ππ β₯ 1.625. Surround suppression is insignificant (ππΌ < 5%) for
most non-ISN solutions.
Page 76
Figure 7. Histogram of the real part of the leading eigenvalue ππΏ
Histogram of the real part of the leading eigenvalue ππΏ for all numerical solutions, color coded
by the minimal suppression index as in Figure 6. Many solutions with strong surround
suppression (ππΌ β₯ 50%) correspond to networks that are close to instability.
Page 78
Top panel: The solid curves represent the excitatory and the inhibitory components of the inverse
connectivity, i.e. πΌ β π (π) β1
ππ
ππ . Parameters in the illustration: πππ = 0.05, πππ = 1.9,
πππ = 0.3, πππ = 0.8, πππ = 0.65, πππ = 0.5, πππ = 0.4 , ππ/ππ = 1.0 and π0 = 0. The
connectivity curves are effectively sharp band pass filters at critical frequency ππΉ . The broken
curves are the Fourier transform of the input functions with different input widths π (in the unit
of 1/ππΉ). The input to the critical frequency (grey dotted line) reaches the maximum when
π = ππΉ β‘ 1/ππΉ . Bottom panel: The solid lines are the excitatory and the inhibitory response
curves of the neurons at the center. Both response curves show strong surround suppression
(ππΌ β₯ 50%), and reach their peak values around critical stimulus size ππΉ , i.e. ππ π~ππΉ and
ππ π~ππΉ.
Page 79
Figure 9. Resonance effects around the critical stimulus size ππΉ
Figure 9a. The ratio of the responses at the critical stimulus size πΈ 0, ππΉ /πΌ 0, ππΉ vs. the ratio
of the output vector 1 + π ππ ππΉ /π ππ ππΉ for all solutions with ππΌ β₯ 50% (286 solutions).
The results show a linear dependency, as predicted in Chapter II, Section 5. The solid line is the
linear least squares fit, given by πΈ/πΌ = 0.87 1 + π ππ /π ππ β 0.08, π2 = 0.94.
Page 80
Figure 9b. Scatter plot of the maximal response stimulus size ππ vs. the critical stimulus size ππΉ ,
for both excitatory and inhibitory populations (different symbols), color coded by suppression
index. Solutions with strong surround suppression (ππΌ β₯ 50%) give ππ ~ππΉ. Solutions with
weaker surround suppression tend to have ππ > ππΉ as predicted in Appendix B, Section 2e.
Page 81
Figure 10. Simulation results of sparse networks
Figure 10a. Population 1 is a sparse network whose corresponding dense matrix has low ππΌ
values in the dense matrix (ππΌπ = 33%, ππΌπ = 21%), simulation parameters: πππ = 0.5, πππ = 1.9,
πππ = 0.3, πππ = 0.65, πππ = 0.4, πππ = 0.5, πππ = 0.4, ππ/ππ = 1.0, π0 = 0 and π = 20%. It
has a uni-modal distribution similar to the experimental results reported by Walker et al (Walker
et al., 2000). Population 2 is another sparse network with stronger surround suppression in the
dense matrix (ππΌπ = 58%, ππΌπ = 48%), simulation parameters: πππ = 0.5, πππ = 1.9, πππ =
0.3, πππ = 0.8, πππ = 0.8, πππ = 0.5, πππ = 0.4, ππ/ππ = 1.0, π0 = 0 and π = 20%. The
distribution is bi-modal, similar to the experimental results with a heavy tail at high ππΌ values
(Jones et al., 2000; Akasaki et al., 2002). Neither example is close to instability, the real part of
Page 82
the leading eigenvalue satisfies ππππ ππΏ < 0.9 for both sparse networks. Furthermore, both
populations contain neurons with very strong surround suppression (ππΌ β₯ 90%).
Page 83
Figure 10b. Simulation results of sparse networks with insignificant population ππΌ value in the
dense matrix (ππΌ = 2%, for both population 3 and 4. Population 3 simulation parameters:
πππ = 0.5, πππ = 1.7, πππ = 0.5, πππ = 0.65, πππ = 0.8, πππ = 0.8, πππ = 3.2, ππ/ππ = 1.0, π0 =
0 and π = 20%. Population 4: πππ = 0.3, πππ = 1.5, πππ = 0.5, πππ = 0.8, πππ = 3.2, πππ =
0.35, πππ = 1.6, ππ/ππ = 1.0, π0 = 0 and π = 20%.). Introduction of sparseness does not produce
strong surround suppression, though these networks are very close to instability: ππππ ππΏ >
0.95
Page 84
Figure 11. Effect of input blurring
Figure 11a. Effect of input blurring in the 3-dimensional π subspace with Gaussian input
functions and blurring width π0 = 0.25, plotted in the same way as Figaure 4b. The distribution
of solutions is similar to the results without input blurring.
Page 85
Figure 11b. Projection of solutions (Gaussian input function and π0 = 0.25) on the πππ vs. πππ
plane. The solid curve is given by πππ2 + πππ
2 = πππ2 = 1. The region above the broken line is
Page 86
given by the biological constraint πππ > πππ . Top panel: color coded by the number of
combinations of other parameters searched. Bottom panel: color coded by ππΌ. Most solutions
satisfy: πππ2 + πππ
2 > πππ2 .
Page 87
Figure 11c. Effect of input blurring in the 3-dimensional π subspace, with Gaussian input
functions and blurring width π0 = 0.25. The solutions cover a larger range for both πΌ β πΈ and
πΌ β πΌ connections. The red markers represent solutions that also appear in the results without
blurring. Such solutions tend to cluster in the local inhibition region where both πΌ β πΈ and πΌ β πΌ
connections are short in range.
Page 88
Figure 11d. Scatter plot of the maximal response stimulus size ππ vs. the critical stimulus size ππΉ
with blurring width π0 = 0.25, plotted in the same way as in Figure 9b. Solutions with strong
surround suppression (dark red) can be roughly divided into two clusters. The cluster along the
diagonal of ππ = ππΉ corresponds to the solutions obtained without input blurring. The other
cluster contains solutions whose maximal response stimulus size ππ is small and is independent
of the critical stimulus size ππΉ .
Page 89
Figure 12. Effects of Rectangular input functions
Figure 12a. 3-dimensional π subspace of the solutions with Rectangular input function and
π0 = 0, plottted in the same way as Figure 4d, color coded by the averaged suppression index of
the excitatory response. Most solutions do not have very strong ππΌ. The solutions with relatively
strong ππΌ have similar distribution in the parameter space compared to solutions with Gaussian
input function.
Page 90
Figure 12b. Same as Figure 12a, color coded by the averaged suppression index of the inhibitory
response.
Page 92
Figure12c. Projection of solutions (Rectangular input function and π0 = 0) on the πππ vs. πππ
plane. The solid curve is given by πππ2 + πππ
2 = πππ2 = 1. The region above the broken line is
given by the biological constraint πππ > πππ . Top panel: color coded by the number of
combinations of other parameters searched. Most of the new solutions have very small πππ .
Bottom panel: color coded by ππΌ. Solutions with significant surround suppression satisfy:
πππ2 + πππ
2 > πππ2 .
Page 93
Figure 12d. 3-dimensional π subspace of the solutions with Rectangular input functions, plotted
in the same way as in Figure 5b, color coded by the averaged suppression index of the excitatory
response. The plane is copied from least squares fit of the solutions with Gaussian input
functions in Figure 5a. Solutions with relatively strong ππΌ have similar distribution to the results
with Gaussian input function.
Page 94
Figure 12e. Same as Figure 12d, color coded by the averaged suppression index of the inhibitory
response. Solutions with relatively strong ππΌ have similar distribution to the results with
Gaussian input function.
Page 96
Figure 13a. Histogram of the critical frequency ππΉ for solutions with Gaussian input and no
blurring. Top panel: all surround suppression solutions (ππΌ > 0%). Most solutions have ππΉ > 0
(7161 out of 7199). Bottom panel: solutions with strong surround suppression (ππΌ β₯ 50%, 286
solutions). All solutions have ππΉ > 0.
Page 97
Figure 13b. Population activity patterns of an example network with various stimulus sizes.
Network parameters: πππ = 0.5, πππ = 1.9, πππ = 0.3, πππ = 0.65, πππ = 0.4, πππ = 0.5,
πππ = 0.4, same as Population 1 in Chapter II Section 6. The critical stimulus size of the network
is ππΉ = 1.1. When the stimulus size is small, the population response is localized around π₯ = 0.
As the stimulus size increases, a population oscillation emerges and becomes strong when the
stimulus size is comparable to the critical stimulus size. When the stimulus size becomes very
large, the population response becomes constant and the oscillation disappears.
Page 98
Figure 14. Spontaneous and sensory-driven activities in the primary visual cortex of ferrets
Page 99
Figure 14. Examples of the spike trains in each age group with all three viewing conditions.
Panel a: P30a (the lower case letter at the end distinguishes multilple sets of experiments done on
the same postnatal day), session 1. Panel b: P44a, session 1. Panel c: P85, session 1. Panel d:
P129, session 1. In the matured age groups P83-P86 and P129-P168, microbursts across all
electrodes can be seen in the spike train.
Page 101
Figure 15a. Top panel: 16 principal components (PC) from age P129, dark viewing condition,
session 1. The π₯ axis in each subplot is the electrode number. The 1st PC is a spatially long-
ranged mode over the 3mm span of the electrode array. Bottom panel: contribution of each PC
mode to the total variance. The 1st PC is a dominant mode.
Page 103
Figure 15b. Spatiotemporal cross-covariance matrix of P129, Dark, session 1. Top panel: the
characteristic structure is approximately a spatially homogeneous mode spanning across all
electrodes with a damped oscillation in the temporal domain. Bottom panel: same plot but with
the 1st PC removed from the spike train. The remaining structure is a short-lived sharp peak
around π = 0, and the cross-covariance is quickly reduced to background level within 50ms.
Page 105
Figure 15c. Top panel: averaged 1st PC mode in each age groups resembles a spatial DC mode
(dotted line: 0.25 on each of the 16 electrodes, i.e. a DC vector of unit length). Bottom panel:
percentage of total variance for all PC modes in different age groups. The 1st PC carries more
variance than the other PC modes; the difference is bigger in matured age groups. Error bars
represent the standard error of the mean.
Page 107
Figure 15d. Temporal correlations of the top 4 PC modes. Top panel: age group P83-P86.
Bottom panel: P129-P168. The 1st PC is a slow temporal mode while the other PCs quickly
decay to the baseline level. The characteristic decay time (defined as the time when the temporal
correlation falls within 2 standard error of the mean from the baseline level) of each PC mode in
each age group is given in Table 2, Chapter III. Error bars represent the standard error of the
mean.
Page 109
Top panel: auto-covariance of first PC in each age group. Upon eye opening P29-P30, the auto-
covariance shows long temporal correlations. This correlation becomes shorter in later age
groups. Oscillation of 8~14Hz is visible in matured age groups. Error bars represent the standard
error of the mean. Bottom panel: example of the curve fitting result in P129, with the model
given by Eq. 11 in Chapter III Section 3. The curve fitting parameters are given in Supplemental
Table 4, Appendix B Section 4b.
Page 111
Top panel: the network oscillation frequency predicted by the connectivity parameters from
surround suppression results in Chapter II, with Gaussian input and no blurring, and the
membrane time constant ππ=10ms. About 85% (6121 out of 7199) of such networks show
oscillations in the spatial DC mode, with a peak around 12Hz (mean Β± std: 12.1 Β± 5.3 Hz). About
37% (2646 out of 7199) of all solutions show oscillations in the 8~14Hz range. Bottom panel:
max ππππ ππ·πΆ vs. ππ»π·πΆ/ππ»3ππ
. The dotted lines are at ππ»π·πΆ/ππ»3ππ
= 5 and ππ»π·πΆ/ππ»3ππ
=
10 respectively, and the highlighted area in between corresponds to the typical ratio of the
characteristic time of the 1st PC vs. the slowest of all other PC modes in matured age groups.
Within this range, 0.825 < max ππππ ππ·πΆ < 0.925; therefore, the Hebbian amplification
predicts the decay time constant to be 57~133ms, in agreement with the decay time constant of
the damped oscillation in age group P129-P168 discussed in Section 3, Chapter III. In addition,
the damped oscillation suggests that Tr π π = 0 = 2 ππππ ππ·πΆ . The damped oscillation
results in age group P129-P168 have Ο1 > 72ππ . LGN inputs typically have correlation time
around 50m. Therefore, the Hebbian amplification contributed by the cortical network gives
ππ» > 22ππ , i.e. ππππ ππ·πΆ > 0.545 if we assume reasonable membrane time constants of 10ms.
Thus, the networks measured in the experiments correspond to inhibition stabilized networks in
our model.
Page 114
Panel a-d: The auto-covariance of the 1st PC mode under Dark, Movie and Noise viewing
conditions for different age groups. Panel a: P29-P30, around eye opening, the temporal
structures of the auto-covariance curves are similar in all 3 viewing conditions, showing long
temporal correlations. Panel b and c: P44-P45 and P83-P86, spontaneous activities have faster
decay time; Movie stimulus significantly increases the temporal correlation. Panel d: P129-P168,
the effect of the Movie input decreases as the animal grows older, resulting in almost identical
auto-covariance curves for Dark and Movie viewing conditions. Oscillations can be observed in
the auto-covariance curves. Especially, strong and sustained oscillation around 10Hz is observed
in the Noise viewing condition. Panel e: Impact of various stimuli for each age group, calculated
as the difference in the auto-covariance curves (as a vector norm) divided by half of the vector
norm of the sum. The differences between spontaneous and sensory-driven activities are small
Page 115
in the eye-opening group P29-P30, and become larger as the animal matures. In late matured
group P129-P168, the difference between Dark and Movie viewing conditions becomes smaller,
but is still larger than that of group P29-P30. The difference between Dark and Noise viewing
conditions becomes larger due to the sustained oscillation induced by Noise stimuli. Error bars in
all panels represent the standard error of the mean.
Page 117
Top panel: Curve fit for P129 Movie with the model given in Equation 11, similar to Figure 16
bottom panel. Curve fit parameters: π1 = 0.439, π1 = 51.3ms, π = 12.6Hz, π0 = 0.23,
π2 = 0.388, π2 = 829ms, π3 = β0.149 and π2 = 0.882. Bottom panel: The power spectra
under all 3 viewing conditions for P129. The bump corresponds to oscillations around 8~14 Hz.
Noise stimuli create strong oscillations with a sharp peak around 10Hz. The second harmonic of
this oscillation can be seen around 20Hz.
Page 120
Log-log plot of the variance vs. the spatial frequency of the Fourier modes. Panel a: P29-P30.
Panel b: P44-P45. Panel c: P83-P86. Panel d: P129-P168. The tuning curve roughly follows a
power law dependency. The dotted lines are the linear least square fits. The fit parameters are
listed in Table 5, Chapter III Section 6. Error bars represent the standard error of the mean.
Page 122
Simulations from a 16-electrode array on a measured orientation map. Top panel: Simulation
results in a model with excitatory neurons only, color coded by different spatial connection
widths: ππ₯ = 0.1, 0.2, 0.5, 1.0, 2.0 mm. Within a biologically reasonable range (ππ₯ > 0.2mm),
the simulation results show a peak around π = 1/750πβ1, corresponding to the typical length
scale for orientation columns. Bottom panel: Simulation results with a deterministic network
containing both excitatory and inhibitory neurons, the connectivity parameters are the same as
Population 1 in Chapter II Section 6: πππ = 1.0, πππ = 0.5, πππ = 1.9, πππ = 0.3, πππ = 0.65,
πππ = 0.4, πππ = 0.5, πππ = 0.4 and ππ = π/8; the dotted line is the linear least squares fit of
the data. The fitting results: π ππππ = β0.200, πππ‘ππππππ‘ = β2.767 and π2 = 0.968. The
simulated data also roughly have a power law dependency over the spatial frequency. The
characteristic length corresponding to orientation columns is absent, similar to the experimental
data but with considerably smaller negative exponent.
Page 123
References
Adini Y., Sagi D. and Tsodyks M. (1997) Excitatory-inhibitory network in the visual cortex:
psychophysical evidence. PNAS, 94(19):10426-10431
Akasaki, T., Sato, H., Yoshimura, Y., Ozeki, H., and Shimegi, S. (2002) Suppressive effects of
receptive field surround on neuronal activity in the cat primary visual cortex. Neurosci. Res. 43,
207β220.
Anderson, J.S., Carandini, M., and Ferster, D. (2000) Orientation tuning of input conductance,
excitation, and inhibition in cat primary visual cortex. J. Neurophysiol. 84, 909β926.
Anderson, J.S., Lampl, I., Gillespie, D.C., and Ferster, D. (2001) Membrane potential and
conductance changes underlying length tuning of cells in cat primary visual cortex. J. Neurosci.
21, 2104β2112.
Angelucci, A., Levitt, J.B., Walton, E.J., Hupe, J.M., Bullier, J., and Lund, J.S. (2002) Circuits
for local and global signal integration in primary visual cortex J. Neurosci. 22, 8633-8646.
Arieli, A., Sterkin, A., Grinvald, A. and Aertsen, A. (1996) Dynamics of ongoing activity:
Explanation of the large variability in evoked cortical responses. Science 273, 1868-1871.
Azouz, R., Gray, C.M., Nowak, L.G. and McCormick, D.A.. (1997) Physiological properties of
inhibitory interneurons in cat striate cortex Cereb Cortex. 7(6): 534-545.
Baker, C. L. and Cynader, M. S. (1986) Spatial receptive-field properties of direction-selective
neurons in cat striate cortex. J. Neurophysiol. 55: 1136-1152.
Berkes, P., Orban, G., Lengyel, M. and Fiser, J. (2011) Spontaneous cortical activity reveals
hallmarks of an optimal internal model of the environment. Science. 331:83-87.
Bonhoeffer, T. and Grinvald, A. (1991) Iso-orientation domains in cat visual cortex are arranged
in pinwheel-like patterns. Nature 353, 429β431.
Buzsaki, G. and Draguhn A. (2004) Neuronal oscillations in cortical networks. Science 304,
1926-1929
Cavanaugh, J.R., Bair, W. and Movshon, J.A. (2002a) Selectivity and spatial distribution of
signals from the receptive field surround in macaque V1 neurons. J Neurophysiol 88, 2547β2556.
Page 124
Cavanaugh, J.R., Bair, W. and Movshon, J.A. (2002b) Nature and interaction of signals from the
receptive field center and surround in macaque V1 neurons. J. Neurophysiol., 88: 2530β2546.
Chapman, B., Stryker, M.P. and Bonhoeffer T (1996) Development of orientation preference
maps in ferret primary visual cortex. J Neurosci 16: 6443β6453.
Chiu, C. and Weliky, M. (2001) Spontaneous activity in developing ferret visual cortex in vivo. J.
Neurosci. 21, 8906-8914.
Chiu, C. and Weliky, M. (2002) Relationship of correlated spontaneous activity to functional
ocular dominance columns in the developing visual cortex. Neuron 35, 1123-1134.
Cooper, N. R., Croft, R. J., Dominey, S. J., Burgess, A. P. and Gruzelier, J. H. (2003) Paradox
lost? Exploring the role of Ξ± oscillations during externally vs. internally directed attention and
the implications for idling and inhibition hypotheses. Int J Psychophysiol. 47, 65-74.
Das, A. and Gilbert, C.D. (1995) Long-range horizontal connections and their role in cortical
reorganization revealed by optical recording of cat primary visual cortex Nature.29; 375(6534):
780-784
DeAngelis, G.C., Freeman, R.D. and Ohzawa, I. (1994) Length and width tuning of neurons in
the catβs primary visual cortex. J. Neurophysiol., 71: 347β374.
Douglas, R.J., Koch, C., Mahowald, M., Martin, K.A., and Suarez, H.H. (1995) Recurrent
excitation in neocortical circuits. Science 269: 981β985.
Dreher, B. (1972) Hypercomplex cells in the catβs striate cortex. Invest. Ophthalmol. Visual Sci.
11: 355-356.
Durack, J.C. and Katz, L.C. (1996) Development of horizontal projections in layer 2/3 of ferret
visual cortex. Cereb Cortex 6:178β183.
Field, D. J. and Tolhurst, D. J. (1986) The structure and symmetry of simple cell receptive-field
profiles in the catβs visual cortex. Proc. R. Soc. Lond. B Biol. Sci. 228: 379-400.
Fiser, J., Chiu, C. and Weliky, M. (2004) Small modulation of ongoing cortical dynamics by
sensory input during natural vision. Nature 431: 573-578.
Fontanini, A. and Katz, DB. (2005) 7 to 12 Hz activity in rat gustatory cortex reflects
disengagement from a fluid self-administration task. J Neurophysiol 93: 2832-2840.
Gilbert, C. D. (1977) Laminar differences in receptive field properties of cells in cat primary
visual cortex. J. Physiol. Lond. 268: 391-421.
Gilbert, C.D. and Wiesel, T.N. (1983) Clustered intrinsic connections in car visual cortex Journal
of Neuroscience 3(5): 1116-1133
Page 125
Goldberg, J.A., Rokni, U., and Sompolinsky, H. (2004). Patterns of ongoing activity and the
functional architecture of the primary visual cortex. Neuron 42: 489β500.
Hammond, P. (1972) Spatial organization of receptive fields of LGN neurons. J Physiol. 222(1):
53-54
Hubel, D. H. and Freeman, D. C., (1977) Projection into the visual field of ocular dominance
columns in macaque monkey. Brain Res. 122: 336β343.
Hubel, D.H. and Wiesel, T.N. (1959) Receptive fields of single neuron in the cat's stirate cortex.
J. Physiol. 148:574-591
Hubel, D.H. and Wiesel, T.N. (1962) Receptive fields, binocular interaction and functional
architecture in the catβs visual cortex. J. Physiol. Lond. 160: 106β154.
Hubel, D.H. and Wiesel, T.N. (1965) Receptive fields and functional architecture in two non-
striate visual areas ( 18 and 19 ) of the cat. J. Neurophysiol. 41: 229-289.
Hubel, D.H. and Wiesel, T.N. (1974) Uniformity of monkey striate cortex: a parallel relationship
between field size, scatter, and magnification factor. J Comp Neurol. 158(3): 295-305.
Jones, H.E., Andolina, I.M., Oakely, N.M., Murphy, P.C. and Sillito, A.M. (2000). Spatial
summation in lateral geniculate nucleus and visual cortex. Exp. Brain Res. 135: 279β284.
Jones, J. P. and Palmer, L. A. (1987a) The two-dimensional spatial structure of simple receptive
fields in the cat striate cortex. J. Neurophysiol. 58: 1187-1211.
Jones, J. P. and Palmer, L. A. (1987b) An evaluation of the two-dimensional Gabor filter model
of simple receptive fields in cat striate cortex. J. Neurophysiol. 58: 1233-1258.
Jones, H. E., Andolina, I. M., Oakely, N. M., Murphy, P. C., and Sillito, A. M. (2000) Spatial
summation in lateral geniculate nucleus and visual cortex Exp. Brain Res. 135: 279-284.
Kato, H., Bishop, P. O., and Orban, G. A. (1978) Hypercomplex and simple/complex cell
classifications in cat striate cortex. J. Neurophysiol. 41:1071-1095.
Kato, H., Bishop, P. O., and Orban, G. A. (1981) Binocular interaction on monocularly
discharged lateral geniculate and striate neurons of the cat. J. Neurophysiol. 46: 932-951.
Katz, L. C. and Shatz, C. J. (1996) Synaptic activity and the construction of cortical circuits.
Science 274: 1133-1138.
Kelly, S. P., Lalor E. C., Reilly R. B. and Foxe J. J. (2006) Increases in Ξ± oscillatory power
reflect an active retinotopic mechanism for distracter suppression during sustained visuospatial
attention. J Neurophysiol. 95: 3844-3851.
Page 126
Kandel E. R. , Schwartz J. H. , Jessell T. M. 2000. Principles of Neural Science, 4th ed.
McGraw-Hill, New York.
Kenet, T., Bibitchkov, D., Tsodyks, M., Grinvald, A. and Arieli, A. (2003) Spontaneously
emerging cortical representations of visual attributes. Nature 425: 954-956.
Kisvarday, Z.F., Toth, E., Rausch, M. and Eysel, U.T. (1997) Orientation-specific relationship
between populations of excitatory and inhibitory lateral connections in the visual cortex of the
cat. Cereb Cortex.7(7): 605-618.
Kuffler S.W. (1953) Discharge patterns and functional organization of mammalian retina. J.
Neuronphysiol. 16:37-68
Leise, E. M. (1990) Modular construction of nervous systems: A basic principle of design for
invertebrates and vertebrates. Brain Res. Rev. 15: 1-23
Leung L. S. (1982) Nonlinear feedback model of neuronal populations in the hippocampal CA1
region. J Neurophysiol 47: 845-868.
Levick, W. R., Cleland, B. G., and Dubin, M. W. (1972) Lateral geniculate neurons of cat:
Retinal inputs and physiology. Investigative Ophthalmology, 11: 302-311.
McCormick, D. A. (1999) Developmental neuroscience β Spontaneous activity: Signal or noise?
Science 285: 541-543.
Meister, M., Wong, R.O.L., Baylor, D.A. and Shatz, C.J. (1991) Synchronous bursts of action
potentials in ganglion cells of the developing mammalian retina. Science 252:939β943.
Movshon, J.A., Thompson, I.D. and Tolhurst, D.J. Receptive field organization of complex cells
in the catβs striate cortex. J Physiol (Lond) 283: 79-99, 1978.
Murphy B. K. and Miller K. D. (2009) Balanced amplification: a new mechanism of selective
amplification of neural activity patterns. Neuron. 61(4):635-48.
Naito, T., Sadakane, O., Okamoto, M., and Sato, H. (2007) Orientation tuning of surround
suppression in lateral geniculate nucleus and primary visual cortex of cat. Neuroscience 149:
962β975.
Nelson, J. I. and Frost, B. J. (1978) Orientation-selective inhibition from beyond the classic
visual receptive field. Brain Res. 139: 359-365.
Ozeki, H., Sadakane, O., Akasaki, T., Naito, T., Shimegi, S., and Sato, H.(2004) Relationship
between excitation and inhibition underlying size tuning and contextual response modulation in
the cat primary visual cortex J. Neurosci. 24: 1428-1438.
Page 127
Ozeki H., Finn I. M., Schaffer E. S., Miller K. D., Ferster D. (2009) Inhibitory stabilization of
the cortical network underlies visual surround suppression. Neuron, 62(4):578-92.
Rao S.C., Toth L.J. and Sur M. (1997) Optically imaged maps of orientation preference in
primary visual cortex of cats and ferrets. J Comp Neurol. 387(3):358-70.
Rose, D. (1977) Responses of single units in cat visual cortex to moving bars as a function of bar
length. J. Physiol. Lond. 27(1): l-23.
Rubin D.B., Zhao J. and Miller K. D. (2009) Spatial periodicity in the V1 extra-classical
receptive field. Swartz Conference 2009.
Salin, P.A., Bullier, J. and Kennedy, H. (1989) Convergence and divergence in the afferent
projections to cat area 17. J Comp Neurol. 283(4):486-512.
Sceniak MP, Hawken MJ and Shapley R. (2001) Visual spatial characterization of macaque V1
neurons. J Neurophysiol. 85(5):1873-87.
Sengpiel F, Sen A, and Blakemore C. (1997) Characteristics of surround inhibition in cat area 17.
Exp Brain Res, 116(2):216-228.
Seung, H.S. (2003). The Handbook of Brain Theory and Neural Networks, Second Edition
(Cambridge, MA: MIT Press), 94-97.
Shu Y. S., Hasenstaub A., Badoual M., Bal T. and McCormick, D. A. (2003) Barrages of
synaptic activity control the gain and sensitivity of cortical neurons. J. Neurosci. 23, 10388-
10401.
Sillito, A. M., Cudeiro, J., and Murphy, P. C. (1993) Orientation sensitive elements in the
corticofugal influence on centre-surround interactions in the dorsal lateral geniculate nucleus.
Exp. Brain Res. 93: 6-16.
Stepanyants A., Martinez, L.M., FerecskΓ³, A.S. and KisvΓ‘rday, Z.F. (2009) The fractions of
short- and long-range connections in the visual cortex. Proc Natl Acad Sci.106(9):3555-3560.
Stettler D.D., Das A., Bennett J. and Gilbert C.D. (2002) Lateral connectivity and contextual
interactions in macaque primary visual cortex. Neuron. 36(4):739-50.
Trefethen, L.N. and Embree, M. (2005) Spectra and pseudospectra: the behavior of nonnormal
matrices and operators. Princeton University Press
Tsodyks M, Skaggs W, Sejnowski T, McNaughton B. (1996) Population dynamics and theta
rhythm phase precession of hippocampal place cell firing: a spiking neuron model. Hippocampus
6:271-280.
Tsodyks, M.V., Skaggs, W.E., Sejnowski, T.J., and McNaughton, B.L. (1997) Paradoxical
effects of external modulation of inhibitory interneurons. J. Neurosci. 17: 4382-4388.
Page 128
Tsodyks, M.V., Kenet, T., Grinvald, A., and Arieli, A. (1999) Linking spontaneous activity of
single cortical neurons and the underlying functional architecture. Science, 286: 1943-1946.
Van Essen, D.C., Newsome, W.T. and Maunsell, J.H. (1984) The visual field representation in
striate cortex of the macaque monkey: asymmetries, anisotropies, and individual variability.
Vision Res., 24: 429-448.
Walker G. A., Ohzawa I., and Freeman R. D. (1999) Asymmetric suppression outside the
classical receptive field of the visual cortex. J Neurosci 19: 10536-10553.
Walker, G.A., Ohzawa, I. and Freeman, R.D. (2000) Suppression outside the classical cortical
receptive field Visual Neuroscience. 17: 369-379.
Ward, L. M. (2003) Synchronous neural oscillations and cognitive processes. Trends in
Cognitive Sciences 7: 553-559.
Webb, B. S., Dhruv, N. T., Solomon, S. G., Tailby, C., and Lennie, P. (2005) Early and late
mechanisms of surround suppression in striate cortex of macaque. The Journal of Neuroscience,
25(50), 11666-11675.
Weliky, M. and Katz, L.C. (1999) Correlational structure of spontaneous neuronal activity in the
developing lateral geniculate nucleus in vivo. Science 285:599-604.
Wilson H. R. and Cowan J. D. (1972) Excitatory and inhibitory interactions in localized
populations of model neurons. Biophys J 12:1-24.
Wolfe, J., and Palmer, L.A. (1998) Temporal diversity in the lateral geniculate nucleus of cat.
Vis. Neurosci. 15: 653-675.
Wong, R.O.L., Chernjavsky, A., Smith, S.J. and Shatz, C.J. (1995) Early functional neural
networks in the developing retina. Nature 374:716-718.
Xue, J. T., Ramoa, A. S., Carney, T., and Freeman, R. D. (1987) Binocular interaction in the
dorsal lateral geniculate nucleus of the cat. Exp. Brain Res. 68: 305-310.
Yamamoto, N., Kurotani, T. and Toyama, K. (1989) Neural connections between the lateral
geniculate nucleus and visual cortex in vitro. Science. 245(4914):192-194.
Page 129
Appendix A: Structures and Functions of the Visual System
The visual system is one of the most studied sub-systems of the central nervous system. The
visual system processes external visual information to build internal representations of the
environment.
1. Cortical and sub-cortical structures in the central visual pathway
In the visual system, visual information flows in a hierarchical order in cortical and sub-cortical
structures along the Central Visual Pathway: visual information is processed in the Retina, the
Lateral Geniculate Nucleus (LGN), the Visual Cortex and the Visual Association Cortex. In sub-
cortical structures, the information processing is strongly sensory-driven. The activity patterns
encode the compressed representations of the feed-forward information. In cortex, image
processing functionalities becomes important; strong recurrent connections (connections within
the same area) and feedback connections from higher areas further refine the complex visual
signals.
The retina is the first structure in the central visual pathway. It contains a large number of
photoreceptor neurons which can be classified into two main cell types named after their
anatomical shapes: the Rod Cells and the Cone Cells. The rod cells respond to monochromatic
stimulus and are sensitive to low luminance. They are responsible for night vision, but in
daylight conditions they are saturated and do not contribute to vision. The cone cells are not
Page 130
sensitive to the lowest luminances but are responsible for vision outside of night conditions,
including color vision. The processing of the visual stimulus starts with the conversion of optical
signals into neuronal signals by the photoreceptors. The resulting spike train is refined in a local
network of Bipolar Cells and Horizontal Cells. In this process, the visual information is
compressed and decorrelated, and is further relayed onto the Ganglion Cells. There are many
more photoreceptors than ganglion cells in the retina. The mapping from photoreceptors to
ganglion cells generally follows a center-surround organization. Depending on the polarity of the
center, the ganglion cells can be classified as on-center cells or off-center cells. On-center cells
receive excitatory connections in the center and inhibitory connections in the surround, while the
off-center cells are just the opposite. Such center-surround organization is responsible for lateral
inhibition in the retina, and is also critical for detection of edges in the visual stimulus.
The processed signals from the retina are sent to the lateral geniculate nucleus (LGN) located in
the thalamus. LGN has the shape of a 'bent knee', and the name 'geniculate' comes from the Latin
word 'genu' for knee. The LGN is the primary relay center between the retina and the cortex.
Neurons in the LGN receive direct inputs from retinal ganglion cells through the Optic Tract;
and neurons in the cortex receive LGN outputs via the Optic Radiation. In primates, the LGN is
generally divided into six layers. Layer 1, 4, 6 receive inputs from the contralateral eye (on the
opposite side), while layer 2, 3, 5 receive inputs from the ipsilateral eye (on the same side).
Layer 1 and 2 contain Magnocellular Cells with large cell bodies and monochromatic response.
Layer 3, 4, 5 and 6 contain Parvocellular Cells with smaller cell bodies and polychromatic
response.
The visual cortex is located in the occipital lobe of the brain and is divided into several visual
areas. On each hemisphere, the primary visual cortex (Visual Area 1, i.e. V1) receives input
Page 131
directly from the ispilateral LGN. The primary visual cortex is divided into six functionally
distinct layers, where layer 1 is the surface layer and layer 6 is the deepest layer. Most of the
LGN inputs are sent to layer 4C, which can be further divided into 2 sub-layers: 4CΞ± and 4CΞ².
Sub-layer 4CΞ± generally receives input from the magnocellular cells in the LGN, while sub-layer
4CΞ² receives mostly parvocellular inputs.
A wide variety of visual features are processed in the primary visual cortex; and the resulting
signals are transmitted into other cortical areas: V2, V3, V4 and V5/MT (Middle Temporal).
Two primary pathways have been identified: the dorsal stream and the ventral stream. The dorsal
stream goes through V2, V5/MT into the posterior parietal cortex. It is associated with
representation of object location and motion, and provides visual information needed to guide
actions such as saccades, reaching, or navigation. As a result, the dorsal stream is usually
referred to as the 'Where Pathway'. The ventral stream goes through V2, V4 into the inferior
temporal cortex. It is associated with object recognition and representation, and therefore is
usually referred to as the "What Pathway". The ventral stream also contributes to the storage of
long-term memory.
2. Receptive field structures of neurons in the visual system.
The receptive field of a neuron is defined as the region in visual space where the presence of a
stimulus will alter the firing of that neuron. In typical receptive field measurements, a
microelectrode is moved very close to the target neuron, and the action potentials are recorded
for various input patterns.
Page 132
As mentioned in the previous section, the receptive field of ganglion cells in the retina consists
of a central disk and a concentric circular surround, which have opposite responses to the
stimulus. There are two types of ganglion cells: on-center cells and off-center cells. An on-center
cell is excited when the center of its receptive field is exposed to light, and is inhibited when the
surround is exposed to light. An off-center cell has just the opposite behavior. The neuronal
response is very strong when the optimal input is presented. When both the center and the
surround areas are stimulated, the responses from both types of cells are very weak. The center-
surround receptive field structure allows ganglion cells to detect stimulus contrast and contour
edges. The size of the receptive field varies for different ganglion cells. Stimuli with low spatial
frequencies are captured by large receptive fields that encode coarse outlines, while the high
spatial frequency signals are detected by small receptive fields. Visual accuracy is the greatest
(0.1 degree) around the fovea where the size of the receptive fields is the smallest. Further down
the central visual pathway, the receptive fields of most LGN cells are similar to those of the
ganglion cells with an antagonistic center-surround structure. The receptive fields of adjacent
neurons in each layer of the LGN correspond to adjacent receptive fields in retina. Neurons in
the same column across layers correspond to the same retinotopic field. In general, most neurons
in sub-cortical areas respond strongly to light spots stimuli. The receptive fields of such neurons
have a simple center-surround structure and are not tuned in orientation.
In the primary visual cortex (V1), except for some neurons in layer 4 with direct LGN input,
most neurons respond optimally to stimulus with more complex spatial features. In particular,
most V1 neurons are selective for the orientation of a light-dark edge, with response on average
diminishing by 50% when orientation differs by 20-30 degrees from the preferred (Hubel and
Wiesel, 1959). The neurons in the primary visual cortex are generally classified as simple or
Page 133
complex cells, based on the spatial structure of their receptive field. The receptive field of a
simple cell is composed of segregated light-preferring and dark-preferring sub-regions, with the
line or lines separating the sub-regions aligned with the cellβs preferred orientation. This
structure can be constructed from the circular receptive fields of LGN cells by, for example,
aligning a row of on-center LGN cells and an adjacent row of off-center LGN along the preferred
orientation. Compared to simple cells, complex cells usually have larger receptive fields.
Complex cells are also orientation-tuned, and respond as if they received inputs from a group of
simple cells of similar preferred orientation but varying locations of their light- and dark-
preferring sub-regions (Hubel and Wiesel, 1962). In general, the receptive field of complex cells
usually cannot be mapped into specific excitatory or inhibitory regions; instead, it demonstrates
some degree of spatial invariance, where preferred stimulus within the receptive field evokes a
response regardless of the exact location.
Typical stimuli used in receptive field measurements are carefully designed drifting gratings,
whose optimal parameters (preferred orientation, spatial and temporal frequency, etc.) are
usually determined in a preliminary search. For simple cells, the response is greatest when the
stimulus pattern matched the excitatory and inhibitory regions in the receptive field, therefore the
drifting grating of optimal orientation will invoke a sinusoidal oscillation (the F1 component) in
the spike train. In contrast, drifting gratings produce a more constant response (the DC
component) in complex cells due to the spatial invariance of the receptive field. To classify the
cell type of a neuron, the F1/DC ratio is calculated from the spike train. A cell is usually
classified as a simple cell if F1/DC > 1.
Stimuli presented in the βcenterβ of the receptive field (the classical receptive field) evoke a
direct response in the target neuron. Stimuli presented in the surrounding areas (the extra-
Page 134
classical receptive field or βsurroundβ) modify the response of the center stimulus, but such
stimuli alone do not evoke a direct response. Most surround stimuli reduce the response of the
center neuron, i.e. the surround suppression effect, as detailed in the introduction in Chapter I,
Section 1.
3. Columnar organization of the visual cortex
In the primary visual cortex of many species, neurons sharing similar response properties are
clustered in functional domains. Such domains extend perpendicular to the surface of the cortex,
i.e. the visual cortex follows a βcolumnarβ organization, in which functional response properties
are invariant vertically (across the layers) and change regularly with horizontal movement across
cortex. The columnar partition of the cortex depends on specific properties of the receptive field.
For example cells with the same eye preference are grouped into ocular dominance columns,
with the eye of dominance varying periodically with horizontal movement across cortex; and
cells with similar preferred orientation are grouped into orientation columns, with the preferred
orientation varying periodically across the cortex. The period of the columns is in the range 0.5-
1.5mm across various species. The columnar organization presumably arises from the local
connectivity circuitry in the cortex: vertical connections across layers and horizontal connections
within 200 microns typically link cells of similar response properties. Such local connections are
much denser than horizontal connections across larger distances. Across different columns, long-
range intrinsic connections tended to link regions of the same ocular dominance and similar
orientation preference.
Page 135
A column is not a discrete structure and does not have precise boundaries. The properties of the
receptive field change gradually across different columns. For example, the cortex encodes a
continuum of preferred orientations (from 0Β° to 180Β°, due to symmetry of the bar-shaped
stimulus), which is typically partitioned by experimenters into 22.5Β° bins to simplify the analysis.
Page 136
Appendix B: Supplemental Information
1. Properties of Inhibition Stabilized Networks
a. Stability of the fixed point
The excitatory and the inhibitory nullclines are given by π
ππ‘πΈ = 0 and
π
ππ‘πΌ = 0, i.e.:
βπΈ + ππ ππππΈ β πππ πΌ + ππ = 0
βπΌ + ππ ππππΈ β ππππΌ + ππ = 0
(Eq. S1)
And the slopes of the nullclines are:
ππΌ
ππΈ πΈ
=1
πππππβ²
(πππππβ² β 1)
ππΌ
ππΈ πΌ
=πππππ
β²
(πππππβ² + 1)
(Eq. S2)
The responses functions ππ and ππ are sigmoid functions, therefore ππβ² > 0, ππ
β² > 0.
For the network given by Equation 1 in Chapter I, consider a linearized dynamics with small
perturbations of πΏπΈ, πΏπΌ around the network fixed point:
π
ππ‘πΏπΈ =
1
ππ
πππππβ² β 1 πΏπΈ β
1
πππππππ
β² πΏπΌ
Page 137
π
ππ‘πΏπΌ =
1
πππππππ
β²πΏπΈ β1
ππ(πππππ
β² + 1)πΏπΌ
(Eq. S3)
The excitatory nullcline (πΏπΌ = 0) is stable if πππππβ² β 1 < 0, i.e. the slope of the excitatory
nullcline is negative. The inhibitory nullcline (πΏπΈ = 0) is stable if πππππβ² + 1 > 0, i.e. the
slope of the inhibitory nullcline is positive. In the ISN model, excitation is unstable, thus
πππππβ² > 1 and the excitatory nullcline has a positive slope.
In a stable network determined by the 2-by-2 connectivity matrix, the trace of the connectivity
matrix must be negative while the determinant must be positive. In the linearized dynamics
around the network fixed point, the connectivity matrix is given by:
π β πΌ = πππππ
β² β 1 βπππππβ²
πππππβ² β πππππ
β² + 1
The stability constraint on the determinant gives:
ππΌ
ππΈ πΈ
= πππππ
β² β 1
πππππβ²
<πππππ
β²
πππππβ² + 1
= ππΌ
ππΈ πΌ
(Eq. S4)
i.e. a necessary condition for stability is that the slope of the inhibitory nullcline must be steeper
than the slope of the excitatory nullcline around the network fixed point.
Page 138
b. Effects of increased inhibitory input
Consider the change in fixed point in response to a small increase πΏππ in inhibitory input in
Equation S1:
βπΏπΈ + πππππβ² πΏπΈ β πππππ
β² πΏπΌ = 0
βπΏπΌ + πππππβ²πΏπΈ β πππππ
β²πΏπΌ + ππβ²πΏππ = 0
This gives:
πΏπΈ =πππππ β²
(πππππ β² β 1)πΏπΌ
πΏπΌ = πππππ
β² β 1 ππβ²
πππππβ² β 1 πππππ
β² + 1 β ππππππππ β²ππ β²πΏππ = β
πππππβ² β 1 ππ
β²
π·ππ‘(π β πΌ)πΏππ
(Eq. S5)
Stability of the network requires that the determinant π·ππ‘ π β πΌ > 0; and for ISN models,
πππππβ² > 1. Therefore, πΏπΈ β πΏπΌ β βπΏππ , so an increase in inhibitory input will lower both the
excitatory and inhibitory activity. For non-ISN models, πππππβ² < 1; thus πΏπΈ β βπΏπΌ β βπΏππ , and
the inhibitory activity will increase while the excitatory activity decreases.
2. Linear rate model with spatial dependency
a. Derivation of the steady state solution
The linear rate model is given by Equation 2 in Chapter II, Section 1:
ππ
ππ‘ πΈ π₯
πΌ π₯ = β
πΈ π₯
πΌ π₯ + π π₯ β²π π₯ β² β π₯
πΈ π₯ β²
πΌ π₯ β² +
ππ π₯
ππ π₯
Page 139
where π is the weight matrix, π₯ can be a scalar or a 2D vector depending on the experiment
setup. For the 2D case, we assume that πππ π₯ β² β π₯ β exp β π₯ β² βπ₯
2
2πππ2 exp β
π¦ β² βπ¦ 2
2πππ2 , i.e. it
is multiplicatively separable.
Applying the convolution theorem on the Fourier transform (π π = ππ₯ π π₯ πβππ βπ₯ , non-
unitary) of Equation 2, we have:
ππ
ππ‘ πΈ(π )
πΌ(π ) = π π β πΌ
πΈ(π )
πΌ(π ) +
π π(π )
π π(π )
At the steady state:
πΈ(π )
πΌ(π ) = πΌ β π π
β1
π π(π )
π π(π )
(Eq. S6)
The inverse Fourier transformation (non-unitary) gives the population activity:
πΈ(π₯ )
πΌ(π₯ ) =
1
2π π ππ πΌ β π π
β1
π π(π )
π π(π ) πβππ βπ₯
(Eq. S7)
where π is the dimension of the system.
In the surround suppression study, we are interested in the response of the neuron at the center of
the stimulus. Thus, for the neuron at π₯ = 0 (Equation 5 in the main text):
Page 140
πΈ 0
πΌ 0 =
1
2π π ππ πΌ β π π
β1
π π π
π π π
b. Steady state solution of 2D model with circular symmetry
With circular stimulus, we study the system in polar coordinates, the weight matrix is:
π ππ = πππππ ππ₯π βπππππππ π π π = 2π ππππ π π½0(πππ)
where π½0(πππ) is the Bessel Function of the first kind.
With Gaussian connectivity, the weight matrix is given by the Hankel transform:
π ππ = 2π πππ ππ₯π βπ2
2π2 π½0 πππ = π2ππ₯π β
π2ππ2
2
(Eq. S8)
In the main text, we assumed general Gaussian connectivity, and simulated response curves with
an isotropic experimental setup. Consequently, the 2D model can be effectively reduced to a 1D
model, expect that in the 2D case the weight matrix depends on π2, while the 1D case depends
on Ο. In the main text, we showed results for the 1D case with π dependency. The results for the
2D model are similar. In the 2D model, due to the π2 dependency, inhibitory connections are
effectively more localized compared to the 1D model. Following the analysis in Chapter II,
Section 3, the network is more likely to function in the ISN regime in the 2D model.
Page 141
c. Analytic solutions with surround suppression boundary conditions
In this sub-section, we simplify the surround suppression constraints on the response curves as
boundary conditions discussed in Chapter II, Section 2.
A'. Summation for small stimulus: the response curve increases with the stimulus size when
the stimulus is small:
π
ππ πΈ(0, π β 0)πΌ(0, π β 0)
> 0
A''. Monotonic suppression for large stimulus: the response curve decreases for large stimulus:
π
ππ πΈ(0, π β β)πΌ(0, π β β)
< 0
B'. Non-negative response for large stimulus:
πΈ(0, π β β)πΌ(0, π β β)
> 0
We solve Equation 5 in the main text for the following Gaussian input function without blurring:
ππ(π₯)ππ(π₯)
= πΆππΆπ
πβ
π₯2
2π2 , and π π(π)
π π(π) =
πΆππΆπ
2πππβ π2π2
2 in Fourier space.
To derive the solutions, we begin with:
π
ππ πΈ(0)πΌ(0)
=1
2π ππ πΌ β π π
β1
πΆππΆπ
π
ππ( 2ππ
β
ββ
πβ π2π2
2 )
=1
2π ππ πΌ β π π
β1
πΆππΆπ
β
ββ
(1 β π2π2)πβ π2π2
2
Page 142
and πΌ β π π β1
= π π π βπ=0
Solutions under constraint A':
We choose a large πβ, so that for any π > πβ:
πΌ β π π β1
β 1 + π ππ₯π βπ2
As π β 0, for any given πβ, we choose a π so that ππβ βͺ 1. Thus for any π < πβ:
1 β π2π2 πβ π2π2
2 β 1 + π π2π2
We define πβ β‘1
πβ:
π
ππ πΈ(0)πΌ(0)
=1
2π ππ πΌ β π π
β1
πΆππΆπ
1πβ
β1πβ
+ π π2
πβ2
+1
2π ππ
πΆππΆπ
(1 β π2π2)πβ π2π2
2
β
1πβ
+1
2π ππ
πΆππΆπ
1 β π2π2 πβ π2π2
2
β1πβ
ββ
+ π ππ₯π β1
πβ2
According to integration by parts:
Page 143
ππ πΆππΆπ
π2π2πβ π2π2
2
β
1πβ
= β π πβ π2π2
2 πΆππΆπ
πβ
1πβ
= βππβ π2π2
2 πΆππΆπ
1πβ
β
+ ππ πΆππΆπ
πβ π2π2
2
β
1πβ
=1
πβπ
β π2
2πβ2 πΆππΆπ
+ ππ πΆππΆπ
πβ π2π2
2
β
1πβ
Therefore:
π
ππ πΈ(0)πΌ(0)
=1
2π ππ πΌ β π π
β1 πΆππΆπ
1πβ
β1πβ
β2
2ππβπ
β π2
2πβ2 πΆππΆπ
+ π π2
πβ2
+ π ππ₯π β1
πβ2
When πβ β β, πβ β 0. Thus in the limit of π β 0:
π
ππ πΈ(0)πΌ(0)
β1
2π ππ πΌ β π π
β1β πΌ
πΆππΆπ
β
ββ
We have:
ππ πΌ β π π β1
β πΌ πΆππΆπ
β
ββ
> 0
(Eq. S9)
Solutions under constraint A'':
In the limit of π β β, let π = ππ,
Page 144
π
ππ πΈ(0)πΌ(0)
=1
2π
1
πππ πΌ β π
π
π
β1
πΆππΆπ
β
ββ
(1 β π2)πβ π2
2
π π/π = π 0 + π΄ π
π
2
+ π( π
π
4
)
where π (0) = π ππ βπ ππ
π ππ βπ ππ
, π΄ = β1
2 π πππππ
2 βπ πππππ2
π πππππ2 βπ πππππ
2 , and π ππ = 2πππππππ , and:
πΌ β π 0 β π΄ π
π
2
β1
= πΌ β π 0 β1
+ πΌ β π 0 β1
π΄ πΌ β π 0 β1
π
π
2
+ π( π
π
4
)
Thus, for Ο β β,
π
ππ πΈ(0)πΌ(0)
β1
2π
1
πππ πΌ β π 0 β π΄
π
π
2
β1
πΆππΆπ
β
ββ
(1 β π2)πβ π2
2
=1
2π
1
πππ πΌ β π 0
β1
πΆππΆπ
β
ββ
1 β π2 πβ π2
2
+1
2π
1
π3ππ πΌ β π 0
β1π΄ πΌ β π 0
β1
πΆππΆπ
β
ββ
π2(1
β π2)πβ π2
2
But
ππβ
ββ
1 β π2 πβ π2
2 = 0
Page 145
While
ππβ
ββ
π2 1 β π2 πβ π2
2 = β π
2
Thus in the limit of π β β
π
ππ πΈ(0)πΌ(0)
~ β1
π3 πΌ β π 0
β1π΄ πΌ β π 0
β1
πΆππΆπ
where
πΌ β π 0 β1
=
1 + π ππ βπ ππ
π ππ 1 β π ππ
π·ππ‘ πΌ β π 0
Since πΌ β π 0 β1
is a 2-by-2 matrix, stability (Constraint C in Section 2 Chapter II) requires
that the determinant π·ππ‘ πΌ β π 0 > 0. Therefore, the condition π
ππ πΈ(0)πΌ(0)
< 0 in the limit of
π β β becomes:
1 + π ππ βπ ππ
π ππ 1 β π ππ
π πππππ
2 βπ πππππ2
π πππππ2 βπ πππππ
2 1 + π ππ βπ ππ
π ππ 1 β π ππ
πΆππΆπ
< 0
(Eq. S10)
Solutions under constraint B':
In the limit of π β β, let π = ππ,
πΈ(0)πΌ(0)
=1
2π ππ πΌ β π
π
π
β1
πΆππΆπ
β
ββ
πβ π2
2
Page 146
π π/π = π 0 + π( π
π
2
)
where π (0) = π ππ βπ ππ
π ππ βπ ππ
. Thus as π β β,
πΈ(0)πΌ(0)
β1
2π ππ πΌ β π 0
β1
πΆππΆπ
β
ββ
πβ π2
2
= πΌ β π 0 β1
πΆππΆπ
πΌ β π 0 β1
=
1 + π ππ βπ ππ
π ππ 1 β π ππ
π·ππ‘ πΌ β π 0
Again, stability requires that π·ππ‘ πΌ β π 0 > 0 , so the requirement of πΈ(0, π β β)πΌ(0, π β β)
> 0
leads to:
1 + π ππ βπ ππ
π ππ 1 β π ππ
πΆππΆπ
> 0
(Eq. S11)
Page 147
d. Expansion of the connectivity filter in the Fourier space when the network
approaches instability
In Equation 5, the denominator of the connectivity filter in the Fourier space πΌ β π π β1
depends on π·ππ‘ πΌ β π π . In most cases, it is difficult to obtain analytic solutions due to the
complexity of this determinant term. When the network approaches instability, π·ππ‘ πΌ β
π π β 0+ (from stability constraint, π·ππ‘ πΌ β π π > 0). We expand the connectivity filter
as a series in π·ππ‘ πΌ β π π by singular value decomposition.
For π΄ = π ππ π
, the singular value decomposition gives: π΄ = πππ+, π΄β1 = ππβ1π+
Where
π =
π2 + π2 + π2 + π2 + π
20
0 π2 + π2 + π2 + π2 β π
2
The left eigenvectors of U (non-normalized): 1
π2+π2βπ2βπ2+π
2(ππ +ππ )
, 1
π2+π2βπ2βπ2βπ
2(ππ+ππ )
The right eigenvectors of V (non-normalized): 1
π2βπ2+π2βπ2+π
2(ππ +ππ)
, 1
π2βπ2+π2βπ2βπ
2(ππ +ππ)
And π = (π2 + π2 + π2 + π2)2 β 4 ππ β ππ 2.
Given π·ππ‘ π΄ = ππ β ππ β 0:
Page 148
π = (π2 + π2 + π2 + π2) 1 β2 ππ β ππ 2
(π2 + π2 + π2 + π2)2+ π ππ β ππ 4
Thus, for the singular value:
π2 + π2 + π2 + π2 β π
2=
(ππ β ππ)2
π2 + π2 + π2 + π2+ π ππ β ππ 4
And for the eigenvectors:
π2 + π2 β π2 β π2 + π
2 ππ + ππ ~
π2 + π2
ππ + ππ=
π2 + π2
π2 ππ
+ π2 ππ
= π2 + π2
(π2 + π2)ππ
β πππ β ππ
π
~π
π
Other expressions of π in the eigenvectors can be simplified in similar ways. Thus we have:
π~
π2 + π2 + π2 + π2 0
0 (ππ β ππ)2
π2 + π2 + π2 + π2
π~1
π2 + π2 π ππ βπ
π~1
π2 + π2 π ππ βπ
It is easy to verify the result:
Page 149
πππ+~π 1π’1π£1+~ π β
ππ β ππ
ππ
π π
~π΄
And the matrix inversion is given by:
π΄β1~1
π 2π£2π’2
+ = π2 + π2 + π2 + π2
ππ β ππ 2 π2 + π2 π2 + π2
πβπ
π βπ
~1
ππ β ππ
π βπ
βπππ
π
~1
ππ β ππ
π βπβπ π
Thus, the connectivity filter can be expressed as:
πΌ β π π β1
~1
π·ππ‘ πΌ β π π
1 + π ππ π
π ππ π 1 + π ππ π βπ ππ π
i.e. Equation 9, in Chapter II Section 5.
e. Relationship between the maximal response stimulus size and the critical
stimulus size
According to Equation 5 in the main text, the response of the center neuron for Gaussian input
function is given by:
πΈ(0)πΌ(0)
=1
2π ππ πΌ β π π
β1
πΆππΆπ
πβ
ββ
πβ π2+π0
2
2π2
Page 150
In this section, we show the derivation for the excitatory neurons. The derivation for the
inhibitory neurons is similar.
We define π π to be the excitatory component of πΌ β π π β1
πΆππΆπ
, around the critical
frequency ππΉ where π·ππ‘ πΌ β π π β 0. By Equation 9 in the main text, π ππΉ ~1/
π·ππ‘ πΌ β π ππΉ . This corresponds to a sharp global peak in the connectivity filter; therefore
we compute the integral using saddle point approximation.
Consider the normalized function π π =πππ π π
πππ π ππΉ , with a distinct global peak π ππΉ = 1. We
examine the excitatory component of the integral:
ππ πΌ β π π β1
πΆππΆπ
πβ
ββ
πβ π2+π0
2
2π2
~ ππ exp πππ π ππΉ π π πβ
ββ
πβ π2+π0
2
2π2
Keeping the second order derivative term in the Taylor expansion of π π , and dropping higher
order terms of 1/π ππΉ as π ππΉ β β, we have
~ ππ π ππΉ exp π β²β² (ππΉ)
2π ππΉ (π β ππΉ)2 π
β
ββ
πβ π2+π0
2
2π2
Note that π β²β² ππΉ < 0 and π ππΉ > 0. We define π β‘ β π β²β² (ππΉ)
2π ππΉ and π β‘
π2+π02
2 to simplify the
notation:
Page 151
ππ π ππΉ exp π β²β² (ππΉ)
2π ππΉ (π β ππΉ)2 π
β
ββ
πβ π2+π0
2
2π2
= ππ ππΉ ππ exp βπ(π β ππΉ)2 β ππ2 β
ββ
= ππ ππΉ ππ exp β(π + π)(π βπ
π + πππΉ)2 β
ππ
π + πππΉ
2 β
ββ
~ππ ππΉ
π + πexp β
ππ
π + πππΉ
2 =ππ ππΉ
π + πexp βπππΉ
2 exp π2ππΉ
2
π + π
Then we examine the dependency of the response on the input size. Since π
πππ = π, we have:
π
ππ
ππ ππΉ
π + πexp βπππΉ
2 ππ₯π π2ππΉ
2
π + π
=π ππΉ ππ₯π βπππΉ
2
(π + π)5/2exp
π2ππΉ2
π + π π + π π +
π02
2 β π2ππΉ
2π2
Thus, at ππΉ = 1/ππΉ:
π
πππΈ(0)
ππΉ
> 0
The same results can be obtained for inhibitory neurons. Combining both results, we have:
π
ππ πΈ(0)πΌ(0)
ππΉ
> 0
Page 152
At ππ , the response reaches its maximum: π
ππ πΈ(0)
πΌ(0)
ππ
= 0. If ππΉ is in the proximity of ππ , then
ππΉ is on the left side of the peak, i.e. ππ > ππΉ .
3. Experimental procedure and data acquisition for spontaneous and sensory-
driven activity in awake ferret V1
The detailed experimental procedure and data acquisition can be found in the paper by Chiu and
Weliky (2001). For the reader's convenience, we list the key points here:
Electrode implant procedure: Briefly, anesthesia was induced and maintained during surgery by
inhalation of isoflurane (0.5β2.0%) in a 2:1 nitrous oxide/oxygen mixture. A section of skull was
exposed over area 17, the dura was reflected and the electrode array aligned along the caudal
bank of the posterior lateral gyrus. After lowering the electrode array so that it touched the
cortical surface, the exposed brain was covered by agar, and a headset was affixed by dental
acrylic to the skull. A separate head-post holder was attached to the fronto-medial skull by
stainless steel screws and dental acrylic. All procedures were approved by the University of
Rochester Committee on Animal Research.
Recording and data acquisition: The multi-electrode array consisted of a single row of 16
electrodes spaced at 200Β΅m (3 mm total span). Each electrode was a 12.5Β΅m-diameter tungsten
wire with 2.5Β΅m H-ML insulation (California FineWire). The insulation along the final 30-60Β΅m
length of wire was removed, creating a 200-300kΞ© impedance electrode. The electrodes
typically provided clear multi-unit signal on each channel with occasional single unit signal. The
average noise amplitude was 5.1Β±0.9Β΅V, whereas the average signal amplitude was 34.4Β±9.8Β΅V.
Page 153
All electrodes were simultaneously raised or lowered by turning a single, 100-thread-per-inch
screw, and they were connected to custom-made amplifiers providing a gain of 20,000. The
signal was band-pass-filtered between 600 and 6,000Hz and digitized at 10,000kHz via an AD
board (National Instruments) to a PC. Data acquisition was performed with custom-written
Labview programs (National Instruments). Spike discrimination was done offline by manually
setting a separate voltage threshold for each electrode. Stable recordings were maintained for 8-
12h. Recordings were initiated after 2-3h of recovery from anesthesia, when the animal was fully
alert. There were delays of approximately 10-20s between interleaved trials.
4. Principal Component Analysis of the activity pattern under Dark, Movie
and Noise viewing conditions
a. The 1st PC mode under Movie and Noise viewing conditions
We perform the same Principal Component Analysis on the recordings with Movie and Noise
inputs as detailed in Chapter III, Section 2. The spatially homogeneous and temporally slow 1st
PC mode also dominates the activity pattern in Movie and Noise viewing conditions, as shown in
Supplemental Figure 1 and 2. Under both view conditions, the 1st PC mode has much slower
characteristic decay time compared to other PC modes, especially in mature age groups, as
shown in Supplemental Table 1 and 2 (mean Β± sem, averaged within each age group, unit in ms).
With movie stimulus:
Supplemental Table 1
Age Group P29-P30 P44-P45 P83-P86 P129-P168
Page 154
PC 1 508 Β± 208 718 Β± 229 523 Β± 186 290 Β± 240
PC 2 197 Β± 95 143 Β± 71 71 Β± 27 49 Β± 13
PC 3 144 Β± 64 114 Β± 58 62 Β± 34 35 Β± 16
PC 4 140 Β± 80 81 Β± 56 50 Β± 33 37 Β± 10
PC 5 112 Β± 77 74 Β± 78 37 Β± 18 28 Β± 7
PC 6 61 Β± 37 77 Β± 97 34 Β± 11 27 Β± 4
PC 7 71 Β± 64 52 Β± 16 31 Β± 11 26 Β± 4
PC 8 55 Β± 32 44 Β± 19 27 Β± 9 26 Β± 3
PC 9 45 Β± 27 51 Β± 16 26 Β± 10 20 Β± 3
PC 10 42 Β± 16 46 Β± 13 23 Β± 8 18 Β± 3
PC 11 37 Β± 15 43 Β± 15 19 Β± 6 16 Β± 3
PC 12 33 Β± 12 39 Β± 12 21 Β± 7 14 Β± 2
PC 13 27 Β± 10 38 Β± 12 19 Β± 5 13 Β± 2
PC 14 23 Β± 6 35 Β± 12 17 Β± 5 13 Β± 2
PC 15 19 Β± 4 29 Β± 8 17 Β± 5 13 Β± 2
PC 16 11 Β± 2 20 Β± 8 16 Β± 5 13 Β± 2
With Noise stimulus:
Supplemental Table 2
Age Group P29-P30 P44-P45 P83-P86 P129-P168
PC 1 486 Β± 208 469 Β± 338 209 Β± 139 116 Β± 133
PC 2 400 Β± 192 129 Β± 154 41 Β± 21 30 Β± 8
PC 3 335 Β± 157 100 Β± 74 46 Β± 38 25 Β± 3
PC 4 248 Β± 126 115 Β± 138 52 Β± 50 25 Β± 2
PC 5 206 Β± 95 95 Β± 118 35 Β± 17 19 Β± 2
PC 6 140 Β± 97 60 Β± 21 31 Β± 20 20 Β± 2
PC 7 121 Β± 77 55 Β± 32 26 Β± 9 20 Β± 2
Page 155
PC 8 93 Β± 61 43 Β± 10 25 Β± 9 19 Β± 3
PC 9 88 Β± 74 50 Β± 9 23 Β± 8 18 Β± 2
PC 10 67 Β± 31 51 Β± 10 22 Β± 8 16 Β± 2
PC 11 60 Β± 28 45 Β± 7 19 Β± 7 15 Β± 2
PC 12 54 Β± 20 43 Β± 6 19 Β± 6 12 Β± 2
PC 13 43 Β± 15 40 Β± 6 19 Β± 6 12 Β± 2
PC 14 33 Β± 12 35 Β± 6 17 Β± 3 12 Β± 2
PC 15 29 Β± 17 30 Β± 7 15 Β± 5 12 Β± 2
PC 16 13 Β± 4 23 Β± 8 14 Β± 4
14 Β± 3
b. Nested model test
In Chapter III Section 3, we proposed a model containing three terms: the damped oscillation
term, the exponential decay term and the constant term (baseline):
πΆ1 exp βπ‘
π1 cos 2πππ‘ + π0 + πΆ2 exp β
π‘
π2 + πΆ3
Eq. 11, Chapter III Section 3
In this sub-section we perform a nested model test to check if the data is over-fitted by additional
parameters. We study 2 nested models within the full model of Equation 11: an exponential
decay model πΆ2 exp βπ‘
π2 + πΆ3 (where πΆ1 = 0), and a baseline model πΆ3 (i.e. a null model
where πΆ1 = πΆ2 = 0). For each day of the experiment, we apply a least squares fit of the data
using all 3 models, and the results are obtained using the Matlab (MathWorks) curve fitting
toolbox with final iteration πΏ < 10β4 and 1000 random seeds per data set. We calculate the
standard error of the mean across all sessions in the data, and compare it to the root mean square
error (RMSE) of each curve fitting model. We perform a Chi-square test with 5% significance
Page 156
level, with the null hypothesis: the RMSE of the tested model is larger than the averaged
standard error of the mean in the data, i.e. the tested model is not a good fit of the data. The
spontaneous activity results in all age groups are listed in the following table:
Supplemental Table 3
Baseline model Exp decay model Full model
P29 Accepted Rejected Rejected
P30a Accepted Rejected Rejected
P30b Accepted Rejected Rejected
P44a Accepted Rejected Rejected
P44b Rejected Rejected Rejected
P45 Accepted Rejected Rejected
P83a Accepted Rejected Rejected
P83b Accepted Accepted Accepted
P85 Accepted Accepted Rejected
P86 Accepted Accepted Accepted
P129 Accepted Accepted Rejected
P134 Accepted Accepted Rejected
P135 Accepted Accepted Rejected
P142 Accepted Rejected Rejected
P151 Rejected Rejected Rejected
P168 Accepted Accepted Rejected
In data sets P29, P30a, P30b (i.e. all of the P29-P30 age group), P44a, P44b, P45 (all of the P44-
P45 age group), P83a, P142 and P151, the null hypothesis is rejected by either the baseline
model or the exponential decay model. Thus, the damped oscillation term is not necessary to
explain the data at 5% significance level. In P83b and P86, none of the models rejects the null
Page 157
hypothesis, i.e. none of them provides a good fit of the data. In P85, P129, P134, P135 and P168,
the null hypothesis is only rejected by the full model with the damped oscillation term, i.e. the
full model provides a good fit of the data, and the damped oscillation term significantly improves
the curve fitting results. The curve fitting parameters from P85 and all ages in age group P129-
P168 are listed in the following table:
Supplemental Table 4
Age P85 P129 P134 P135 P142* P151* P168
ππ 0.729 0.168 0.145 0.153 0.133 0.120 0.119
ππ(ms) 26.2 77.2 96.0 94.1 35.7 82.3 72.1
π (Hz) 12.2 13.6 13.8 8.57 16.9 9.09 10.7
ππ 1.97 0.174 0.551 2.50 1.48 2.07 1.85
ππ 0.146 0.282 0.073 0.059 0.241 2.58 0.073
ππ(ms) 131 326 1015 221 602 2709 347
ππ 0.072 0.000 0.000 0.064 0.000 0.000 0.000
ππ 0.779 0.878 0.863 0.544 0.346 0.473 0.577
* A similar damped oscillation can also be seen in the curving fitting results from P142 and P151,
though not significant.
Page 158
5. Mechanisms of Hebbian amplification and properties of normal and non-
normal matrices
a. Hebbian amplification for translation-invariant linear rate models
In this section, we study a linear rate model:
π0
π
ππ‘π (π‘) = π β πΌ π (π‘) + π (π‘)
Where π is the connectivity matrix, πΌ is the identity matrix and π π‘ is the external input.
Assuming the input started at a distant past, the solution is given by:
π (π‘) = ππ‘ β²πΎ π‘ β π‘ β² π π‘β² π‘
ββ
where πΎ π‘ β π‘β² = exp 1
π0 π β πΌ π‘ β π‘β² is the temporal kernel.
Next we choose a set of orthogonal bases π = πππ π , and the firing rate of the i-th mode is:
ππ(π‘) = ππ‘ β²πΎππ π‘ β π‘ β² ππ π‘β²
π‘
ββπ
Thus, the auto-covariance of the i-th mode at time lag π β₯ 0 is given by:
ππ π‘ ππ π‘ + π π‘
= ππ‘ β²π‘
ββππ
ππ‘ β²β² πΎππ π‘ β π‘ β² ππ π‘β²
π‘+π
ββ
πΎππ π‘ + π β π‘ β²β² ππ π‘ β²β² π‘
Page 159
= ππβ
0ππ
πππΎππ π πΎππ π β
0
ππ π‘ β π ππ π‘ β π + π π‘
where π β‘ π‘ β π‘β², π β‘ π‘ + π β π‘β²β² and π(π‘) π‘ is the average of π(π‘) over time.
For white noise stimulus: ππ π‘ β π ππ π‘ β π + π π‘ = πΏππ πΏ(π β π + π)
ππ π‘ ππ π‘ + π π‘ = πππΎππ π πΎππ π + π β
0π
In general the temporal kernel πΎ π is a matrix exponential, and the analytic solutions can be
difficult to obtain. When the connectivity matrix is translation-invariant, in a model with a single
population, the temporal kernel will diagonalize over the Fourier bases. For models with both
excitatory and inhibitory populations, the connectivity matrix contains 4 sub-matrices, e.g.
Equation 3, Chapter II. The temporal kernel diagonalizes within each sub-matrix, and the
dynamics at a given spatial frequency π is determined by a 2-by-2 matrix. When the eigenvalues
dictate the dynamics of the network, we have πΎππ π β πΏππ exp ππ β 1 π , where ππ is the
eigenvalue of the corresponding Fourier mode in the connectivity matrix. Therefore, the auto-
covariance is given by:
ππ π‘ ππ π‘ + π π‘ β1
2 1 β ππ exp ππ β 1 π
(Eq. S12)
b. Properties of normal and non-normal matrices
A matrix π is normal if it commutes with its conjugate transpose: πβ π = ππβ . Normal
matrices can be diagonalized by a unitary transform, and the dynamics of normal matrices are
Page 160
determined by the eigenvalues of the orthogonal bases. In contrast, non-normal matrices cannot
be diagonalized by a unitary transform, and their eigenvectors are not mutually orthogonal.
The normality of a matrix can be characterized by π = π πβ1 , where the norm β is
usually chosen as the matrix 2-norm. For normal matrices, π = 1; while π > 1 for non-normal
matrices. A common tool to study non-normal matrices is the Ξ΅-pseudospectrum on the complex
plane. The Ξ΅-pseudospectrum of a matrix π is defined as:
πν π = π§: (π§πΌ β π)β1 > νβ1, π§ β β
For normal matrices, the pseudospectra πν π are concentric disks of radius ν about each
eigenvalue ππ , i.e. the spectra of the matrices, denoted as πν(ππ). For highly non-normal matrices
(π β« 1), the pseudospectra πν π can be very different from πν(ππ) even at very small Ξ΅.
Consequently, non-normal matrices demonstrate transient dynamics that are not characterized by
the eigenvectors. However, such dynamics asymptote to the ones determined by the eigenvalues
as π‘ β β (Trefethen and Embree, 2005).
Page 162
Supplemental Figure 1. PCA results of Movie viewing conditions
PCA results of Movie viewing conditions, plotted in the same way as Figure 15c.
Top Related